Artificial Intelligence Review最新文献

In recent years, deep learning algorithms have been widely regarded as the preferred method for remote sensing image analysis and have been successfully applied in the field of change detection. However, currently only a few comprehensive review papers on deep learning based remote sensing image change detection methods are available. To address the above issues, this study provides the current research status and development trends of remote sensing image change detection and analysis based on deep learning. Firstly, considering the differences in data volume and in data characteristics between different data sources, unlike previous reviews that only focus on the change detection problem of a certain type of remote sensing data, this review outlines the various data types involved in remote sensing image change detection, mainly including very high-resolution data, hyperspectral data, synthetic aperture radar data, and heterogeneous data. Secondly, unlike previous reviews that only introduce deep learning methods as a category of methods, this review comprehensively summarizes the research progress in remote sensing image change detection from three aspects: supervised deep learning, semi-supervised deep learning, and unsupervised deep learning, and explores the advantages and limitations of these methods. On this basis, we also propose five interesting research directions to promote further development in this field, including data privacy protection, multi-modality semantic-level change detection, lightweight models, change detection assisted by foundational models, as well as brain-inspired and vision-language models for change detection. This study will help deepen our understanding of deep learning in change detection in multiple ways and lay the foundation for future research.

The zeroing neural network (ZNN) possesses core characteristics such as a simple architecture and fast response speed. These advantages have earned it widespread research attention and spurred the generation of high-quality outcomes in related studies. This paper presents a comprehensive introduction to ZNN. It begins with an elaboration on the core principles of ZNN, clarifying its underlying operational mechanisms from a theoretical perspective. The evolution of ZNN is discussed, and an in-depth analysis is conducted on continuous-time ZNN (CTZNN) and discrete-time ZNN (DTZNN) respectively. Subsequently, the paper systematically reviews and summarizes ZNN’s application practices in various fields, including mathematical applications, robot control, and image processing. Finally, this paper summarizes the current challenges faced by ZNN, proposes targeted future research directions, and provides new perspectives for researchers to understand and apply ZNN.

Human cognition falters under overload because working memory is sharply limited, as described by Cognitive Load Theory. Advanced AI systems show parallel failures when tasks exceed context windows or cause model collapse. This review synthesizes these constraints through a unifying lens, revealing shared mechanisms like bounded workspaces and chunking, alongside divergences such as human metacognition. We introduce a “bounded agent complementarity” model that proposes dynamic load-balancing for symbiotic intelligence, with implications for reasoning in domains such as education, medicine, and aviation. The framework highlights ways to mitigate these mutual limits and yields testable predictions for augmented cognition and resilient human-AI systems.

Over the past decade, recommender systems have become pivotal across digital platforms, supporting tasks such as media choice, e-commerce navigation, industrial decision support, and personalized learning. By analyzing user behaviors and preferences, modern engines enable filtering, ranking, and adaptive interaction at scale. A recent research trend concerns situation-aware recommenders, that are systems able to perceive and interpret surrounding conditions to adapt their output and anticipate user goals. These systems are increasingly shaped by the need for transparency, reliability, and alignment with trustworthy AI principles. Despite growing interest, the literature lacks a clear conceptual definition of “situation”, a distinction from “context”, and unified models, design guidelines, and evaluation frameworks for truly situation-aware recommenders. Consequently, only a limited subset of deployed solutions integrates situation awareness in an explicit and systematic way. This work presents a systematic review and classification of modern situation-aware recommender systems, highlighting the most used techniques, domains of application, open issues, and research challenges. The review follows the PRISMA methodology for systematic literature studies. The analysis is completed with the proposal of a reference framework grounded in Endsley’s three-level Situation Awareness model and aligned with emerging principles of trustworthy AI. The architecture is used to identify key challenges and outline research directions in the field. It also serves as a comparative lens for existing work and as a blueprint intended to guide the development of the next generation of transparent, reliable, and human-centered recommender systems.

Deep neural networks such as convolutional neural networks (CNNs) and transformers have achieved many successes in image classification in recent years. It has been consistently demonstrated that best practice for image classification is when large deep models can be trained on abundant labeled data. However, there are many real world scenarios where the requirement for large amounts of training data to get the best performance cannot be met. In these scenarios, transfer learning can help improve performance. To date, there have been no surveys that comprehensively review deep transfer learning as it relates to image classification overall. We believe it is important for future progress in the field that current knowledge is collated and the overarching patterns analyzed and discussed. In this survey we formally define deep transfer learning and the problem it attempts to solve in relation to image classification. We survey the current state of the field and identify where recent progress has been made. We show where the gaps in current knowledge are and make suggestions for how to progress the field to fill in these knowledge gaps. We present new taxonomies of the solution and applications of transfer learning for image classification. These taxonomies make it easier to see overarching patterns of where transfer learning has been effective and, where it has failed to fulfil its potential. This also allows us to suggest where the problems lie and how it could be used more effectively. We demonstrate that under this new taxonomy, many of the applications where transfer learning has been shown to be ineffective or even hinder performance are to be expected when taking into account the source and target datasets and the techniques used. In many of these cases, the key problem is that methods and hyperparameter settings designed for large and very similar target datasets are used for smaller and much less similar target datasets. We identify alternative choices that could lead to better outcomes.

The automotive sector is undergoing continuous technological evolution driven by the demand for sustainable and safe vehicles. Among the main factors influencing safety, driver behaviour has been identified as a critical contributor to road crashes. This systematic review explores recent innovations in detecting risky driver behaviours, addressing six research questions: the most relevant datasets used for algorithm development and evaluation; system architectures and methodologies for anomaly detection; the most studied driver behaviours and related environmental, human, and mechanical factors; advances in machine learning, deep learning, and statistical methods; performance metrics and validation approaches; and the role of embedded technologies and sensors in practical applications. The review included 93 peer-reviewed articles published between 2020 and 2024, sourced from ACM, IEEE, ScienceDirect, and Scopus. Exclusion criteria were duplicates, non-open access, retracted works, and studies unrelated to outlier detection or driver behaviour. The Parsifal tool was used to support systematic data processing. Results highlight the most frequently used datasets, proposed models, and their performance in detecting driver behaviours, as well as the influence of contextual factors such as traffic rules, road conditions, and sensor limitations. Despite advances, real-world integration remains challenging, requiring further research and development. This review aims to guide researchers in understanding the current state of anomaly detection in driving contexts and to emphasize the need for broader collaboration to create effective, deployable solutions that enhance road safety worldwide.

Deep clustering has been dominated by flat models that split a dataset into a predefined number of groups. Although recent methods achieve an extremely high similarity to the ground truth on popular benchmarks, the information contained in a flat partition is limited. In this paper, we introduce CoHiClust, a Contrastive Hierarchical Clustering model based on deep neural networks, which can be applied to typical image data. By employing a self-supervised learning approach, CoHiClust distills the base network into a binary tree without access to any labeled data. The hierarchical clustering structure can be used to analyze the relationship between clusters, as well as to measure the similarity between data points. In addition to the hierarchical structure we propose two visualization techniques, which allow us to deliver an intuitive explanation of tree nodes. Experiments demonstrate that CoHiClust generates a reasonable structure of clusters, which is consistent with our intuition and image semantics. Moreover, it obtains superior clustering accuracy on most of the image datasets compared to the state-of-the-art flat clustering models.

Federated learning is a promising method for developing collaborative and privacy-preserving medical AI models in dispersed geographical areas. While FL enables the development of robust and generalizable imaging models by leveraging diverse data sources distributed across different institutions, it has not yet been fully standardized, and numerous open issues remain to be resolved. These include interoperability, security vulnerabilities, optimal strategies for model aggregation, and privacy safeguarding. In this critical review, a thorough survey of FL methods on MRI image analysis has been performed. In particular, parameters such as the models used, datasets, harmonization techniques, aggregation methods, security and privacy measures, “real-world” implementations, and performance measurements were investigated. This review sheds light, from a critical perspective, on the required infrastructure of developing deep federated learning applications based on MRI data such as lesion detection, classification, image reconstruction, and tumor segmentation. Unlike previous FL reviews, this work’s main focus is MRI since its tomographic data structure, acquisition protocol variability, and inter-scanner heterogeneity introduce challenges not typically present in modalities such as ultrasound, X-ray, or computed tomography. At the same time, the wide use of MRI in clinical routine makes it an excellent fit for FL architectures that leverage diverse and heterogeneous data sources to develop generalized models while safeguarding patient confidentiality. The present review addresses an important gap in the literature by offering domain-specific insights needed to advance the development of reliable MRI-based FL systems.

The Cognitive and Artificial Intelligence Evaluation (CAIE) framework provides a structured and domain-independent methodology for assessing the intelligence of artificial and information systems in a broader perspective. The primary achievement of this research is the categorization of over ninety cognitive features into six evaluation zones, supported by a two-stage scoring model that combines detailed feature-level analysis with higher-level structural interpretation. This approach has proven effective in identifying system maturity and developmental potential, offering systematic insights into both strengths and weaknesses across cognitive domains. The practical validation through use-case analysis demonstrates that CAIE is adaptable to diverse technological contexts, enabling consistent comparison between AI and non-AI systems. By treating cognitive features as measurable and comparable attributes, the framework introduces a coherent mechanism for benchmarking, scalability, and strategic development. The main contribution of this work lies in advancing both academic research and real-world implementation by delivering a cognitively informed, practically relevant tool that bridges theoretical evaluation concepts with actionable methods for designing and improving intelligent systems.

In edge computing environments using a serverless approach, properly managing resources is essential to reduce energy usage and maintain continuous service for IoT devices. Because the energy availability of edge nodes fluctuates over time, scheduling tasks in real-time becomes complex, requiring the system to allocate resources dynamically while still responding promptly to incoming requests. In this paper, we propose autonomous energy-aware scheduler mechanism (AEASM), a novel scheduling framework that selects and executes the most suitable scheduler based on the predicted energy levels of active nodes. AEASM leverages a MAPE-K control loop to continuously monitor energy states, analyze workload requirements, plan scheduling decisions, and execute the chosen scheduler. Additionally, AEASM enhances network availability by dynamically adapting to energy drops in active nodes. Our evaluation across four workload distribution models demonstrates that AEASM optimizes energy consumption, ensures sufficient resource allocation for incoming requests, increases network availability for over one hour, and improves overall quality of service through fault tolerance and faster response times.