Artificial Intelligence Review最新文献

Car theft has become a significant issue in modern societies, with far-reaching individual and social consequences. This criminal act causes substantial financial losses for vehicle owners, undermines public trust in security systems, and increases social and governmental costs. Therefore, research on developing innovative and efficient methods for detecting and preventing car theft holds particular importance. In this study, advanced methods for detecting car theft have been evaluated and compared through two main approaches: deep learning and machine learning. First, pre-trained deep neural networks were examined. In the second phase, various image features were extracted using feature extraction methods, such as Edge Direction Histogram (EDH), Edge Oriented Histogram (EOH), and Histogram Oriented Gradient (HOG), followed by the assessment of machine learning approaches. Finally, a hybrid model based on Hybrid Edge and Gradient-Based Features (HFEM) combined with an XGBoost classifier was proposed, achieving an accuracy of 98.6% in predicting car theft.

Facial expression recognition (FER) remains a challenging task due to subtle variations in facial details and unconstrained conditions such as changes in head posture, illumination, and occlusion. Current FER approaches primarily focus on capturing discriminative facial features in vision manner, often neglecting the rich semantic information available in textual modalities. Additionally, these methods typically rely on generic classification templates, which fail to capture instance-specific features, resulting in inadequate representation and fine-grained discrimination ability. To tackle the above issues, we propose a novel emotion-aware adaptation framework that integrates the pre-trained CLIP model for FER, leveraging both visual and textual modalities to enhance representation learning and capture fine-grained emotional details. Specifically, we introduce the Expression-aware adapter module to capture emotion-specific facial representations through task-specific fine-tuning while preserving the generalization capabilities of the CLIP model. Furthermore, the instance-enhanced expression classifier module is proposed to enhance textual descriptors with instance-specific visual embeddings using spherical linear interpolation, creating a more precise and discriminative classifier. Extensive experiments on three in-the-wild FER benchmarks demonstrate superiority of our proposed approach.

Bipedal robots are gaining global recognition due to their potential applications and the rapid advancements in artificial intelligence, particularly through deep reinforcement learning (DRL). While DRL has significantly advanced bipedal locomotion, the development of a unified framework capable of handling a wide range of tasks remains an ongoing challenge. This survey systematically categorises, compares, and analyses existing DRL frameworks for bipedal locomotion, organising them into end-to-end and hierarchical control schemes. End-to-end frameworks are evaluated based on their learning approaches, whereas hierarchical frameworks are examined in terms of their layered structures that integrate learning-based and traditional model-based methods. We provide a detailed evaluation of the composition, strengths, limitations, and capabilities of each framework. Furthermore, this survey identifies key research gaps and proposes future directions aimed at creating a more integrated and efficient unified framework for bipedal locomotion, with broad applicability in real-world environments.

The rapid development of Transformer-based large language models (LLMs) has made them one of the most critical technological infrastructures in modern society. However, this rapid deployment has transformed the risk of privacy breaches from a theoretical concern into a systemic threat spanning the entire lifecycle of LLMs. These risks continually challenge existing data compliance and regulatory frameworks, directly limiting the large-scale adoption of LLMs in highly sensitive and heavily regulated industries. Cryptographic technologies, such as fully homomorphic encryption (FHE) and secure multi-party computation (MPC), have garnered significant attention due to their provable security guarantees, theoretically safeguarding the privacy of sensitive data and LLMs weights. These cryptographic techniques have rapidly permeated key stages of LLMs, including data selection, fine-tuning, and inference. Despite these advancements, there is currently no comprehensive survey summarizing the work related to cryptography-based privacy-preserving LLMs (CPLMs), leaving their research isolated and fragmented. To fill this gap, We provide a comprehensive review of existing CPLMs research and systematically classifies them, enabling researchers to effectively coordinate optimization strategies for the efficient design of CPLMs algorithms. Finally, based on the limitations of current CPLMs research, we outline several promising directions for future exploration.

Patients increasingly rely on easily accessible online resources, often ignoring source credibility. Large Language Models such as ChatGPT and DeepSeek provide free, near human interaction on any imaginable topic, including medical conditions. While the benefits provided by this technology are evident and undeniable, concerns regarding the reliability and safety remain. In this study, we assessed the quality, safety, and reproducibility of responses generated by ChatGPT-4o mini and DeepSeek-R1 on the urolithiasis - an increasingly prevalent condition with complex aetiology and diverse management options. We screened for the most frequently asked questions on kidney stone disease. A set of 76 questions was generated and divided into six categories: general information, risk factors, symptoms, diagnosis, treatment and prognosis. Each question was entered into DeepSeek-R1 and ChatGPT-4o mini. Responses were independently evaluated by two attending urologists using a four-point scale based on clearly defined, pre-established criteria. Discrepancies were resolved by a third expert. Cosine similarity index was applied to evaluate the degree to which LLM responses remained stable over time in wording and meaning. Direct comparisons on the response lengths were conducted. Initial analysis with no category differentiation favoured DeepSeek R1 (p < 0.001). The worst outcomes for both models were recorded in the “treatment” category, yet with DeepSeek’s statistically significant advantage. Moreover, the Chinese LLM provided more accurate responses in “general information” category. The median cosine similarity score for responses generated by DeepSeek-R1 and ChatGPT-4o was 0.7 (IQR 0.655–0.736) and 0.86 (IQR 0.805–0.9), respectively. Responses from DeepSeek-R1 were significantly shorter, with a median word count of 385.5 (330.5–448.5) compared to and 672.5 (438–873.25) words for ChatGPT-4o mini (p < 0.001). Additionally, DeepSeek-R1 responses were more consistent in terms of length exhibiting a narrower distribution when compared to ChatGPT-4o mini. Among the evaluated LLMs available free of charge, DeepSeek-R1 emerged as a more accurate and concise source of patient information, while ChatGPT-4o mini demonstrated significantly greater reproducible responses. The reasoning process of DeepSeek-R1 has the potential to enhance patient comprehension of complex medical concepts thereby improving treatment adherence. Nevertheless, limitations of LLMs such as susceptibility to hallucinations and biases derived from their training data must be carefully considered.

The proliferation of metaheuristic optimization algorithms has led to concerns about their novelty. This study introduces three key contributions to address this challenge: (1) a novel systematic taxonomic framework that employs nineteen rigorously selected, metaphor-free criteria to evaluate algorithmic distinctiveness; (2) a comprehensive clustering methodology that combines Rogers-Tanimoto distance analysis with principal component analysis (PCA) and hierarchical clustering to quantify algorithmic similarities; and (3) an objective assessment method for evaluating genuine algorithmic innovations. Through the analysis of 145 metaheuristic algorithms, we demonstrate that 74 algorithms (51.0%) exhibit distances below the confidence interval threshold, indicating profound structural overlap. Network analysis reveals 26 algorithms with perfect structural identity (distance = 0.0) and 512 algorithm pairs showing high similarity (distance < 0.039), representing 18.9% of all pairwise comparisons. The results show that numerous algorithms claiming innovation deliver only incremental modifications to existing implementation patterns, lacking fundamental methodological advancement. The framework provides both a theoretical foundation for understanding algorithmic similarities and a practical tool for evaluating new algorithmic proposals, potentially transforming how the field assesses and develops novel optimization methods.

Topological data analysis (TDA) is a rapidly evolving field in applied mathematics and data science that leverages tools from topology to uncover robust, shape-driven, and explainable insights in complex datasets. The main workhorse is persistent homology, a technique rooted in algebraic topology. Paired with topological deep learning (TDL) or topological machine learning, persistent homology has achieved tremendous success in a wide variety of applications in science, engineering, medicine, and industry. However, persistent homology has many limitations due to its high-level abstraction, insensitivity to non-topological changes, and restriction to point cloud data. This paper presents a comprehensive review of TDA and TDL beyond persistent homology. It analyzes how persistent topological Laplacians and Dirac operators provide spectral representations to capture both topological invariants and homotopic evolution. Other formulations are presented in terms of sheaf theory, Mayer topology, and interaction topology. For data on differentiable manifolds, techniques rooted in differential topology, such as persistent de Rham cohomology, persistent Hodge Laplacian, and Hodge decomposition, are reviewed. For one-dimensional (1D) curves embedded in 3-space, approaches from geometric topology are discussed, including multiscale Gauss-link integrals, persistent Jones polynomials, and persistent Khovanov homology. This paper further discusses the appropriate selection of topological tools for different input data, such as point clouds, sequential data, data on manifolds, curves embedded in 3-space, and data with additional non-geometric information. A review is also given of various topological representations, software packages, and machine learning vectorizations. Finally, this review ends with concluding remarks.

Convolutional Neural Networks (CNNs) have rapidly transformed the landscape of medical image analysis, particularly in the detection and management of diabetic retinopathy (DR), a leading cause of irreversible vision loss worldwide. Since their inception in early architectures such as LeNet and AlexNet, CNNs have evolved into advanced designs including VGGNet, Inception, ResNet, U‑Net, DenseNet, EfficientNet, and ConvNeXt, enabling increasingly precise classification, detection, and segmentation of retinal lesions. This systematic review synthesizes current evidence on CNN applications across diverse DR tasks, including disease classification, lesion localization, vessel segmentation, diabetic macular edema detection, ischemia assessment, and disease monitoring. Comparative analyses highlight the performance of CNNs relative to traditional machine learning methods, as well as the benefits of ensemble strategies and customized architectures tailored to multimodal imaging. Despite remarkable progress, real‑world deployment faces persistent challenges related to data quality, computational demands, interpretability, clinical reliability, and ethical considerations. Emerging solutions such as lightweight CNNs, hybrid models, and ensemble learning approaches aim to enhance scalability, efficiency, and clinical integration. The findings reveal both the transformative potential and the limitations of CNNs in DR management, underscoring the need for future research on robust, interpretable, and ethically aligned systems. This review provides ophthalmologists, data scientists, and healthcare practitioners with the latest insights into CNN‑based diabetic retinopathy analysis, fostering the advancement of accessible and clinically reliable diagnostic technologies poised to transform patient care and outcomes.

Aspect-level sentiment analysis is crucial for consumers and institutions, enabling them to monitor satisfaction regarding specific aspects of products and services through user reviews. Over time, various artificial intelligence techniques have been implemented with significant success. However, most of these techniques rely heavily on a substantial amount of labeled data. In this context, Cross-Domain Aspect-Based Sentiment Analysis emerges, leveraging data from source domains to enhance performance in the target domain. This systematic review contributes to this framework by outlining the primary solutions developed to tackle this challenge. It presents their data sources, compared methods, and the evolution of the main technologies adopted while identifying gaps that may inspire future research endeavors. A new classification of models is proposed here, considering the cross-domain approach. This fresh perspective aims to assist researchers in their quest for innovation, clarifying the context of their proposal and suggesting relevant comparisons with existing works.

Traffic network delays seriously affect the air transportation system’s safety, economy, and efficiency, and have always been a global concern. Flight delays usually propagate within airport networks, causing subsequent flights to be delayed. However, existing works lack in considering network causality, and the incorporation of emerging large language models (LLMs). Thus, this paper endeavours to examine the literature on network delay prediction that combines different background knowledge with journal paper publishing data. Particularly, the network delay prediction methods are categorized into four aspects: classic methods without explicit network topology modelling, traditional explicit network-based prediction methods, emerging deep learning methods, and the application of LLMs in transportation. Classic methods without explicit network topology modelling, including statistical analysis, operations research, traditional machine learning and causal inference without network structures, offer interpretable baselines but fail to capture the complexity and nonlinearity of air traffic systems. Traditional explicit network-based prediction methods often approach air traffic systems through frameworks such as complex networks and queuing theory, with an increasing focus on causal relationship analysis. However, these methods fall short in capturing the spatiotemporal dependencies of network delays, particularly in modelling spatiotemporal causality. In contrast, emerging deep learning methods have advanced significantly, enabling the construction of spatiotemporal causal networks and improving the accuracy of network delay prediction. In addition, some future trends are analyzed. It is concluded that graph neural networks with causality and emerging deep learning methods (e.g., spatiotemporal GCN) are identified as essential directions. Moreover, a conceptual AirTraffic LLM is suggested via a novel Spatial-Temporal Causal Large Language Model (STC-LLM) framework for high-precision flight delay prediction, which requires further experimental validation and real-world testing. Nevertheless, issues such as data privacy, model opacity, and high computational costs must be carefully addressed when applying LLMs. Finally, the findings are expected to enhance understanding of delay propagation among researchers, practitioners, and policymakers, while providing insights and guidance to airports, airlines, and air traffic control.