Artificial Intelligence Review最新文献

Super-resolution (SR) imaging is a key task in computer vision, and recent progress has been driven by deep learning. However, manually designed SR networks often suffer from poor generalization, inefficiency, and long development cycles. Neural Architecture Search (NAS) offers an automated paradigm to overcome these limitations. However, its application to SR remains in a nascent stage, presenting significant research gaps such as prohibitive computational costs and the limited generalization of searched architectures. This review summarizes advances of NAS in SR, analyzing its essential components search space, search strategy, and performance evaluation and discussing applications in single image SR, remote sensing, and video SR. Studies show that NAS-based models can achieve competitive or superior performance with lower computational cost compared to handcrafted designs. Specifically, we emphasize the following contributions: (1) a comprehensive analysis of NAS components tailored to SR tasks; (2) a review of NAS applications across various SR domains with demonstrated improvements in performance and efficiency; and (3) identification of these unresolved challenges to outline actionable future directions, including reducing search costs, enhancing cross-domain robustness of lightweight models, and expanding NAS applications in SR-related tasks. This work aims to provide theoretical and methodological insights to support research and practical deployment of NAS in SR imaging.

Structured sentiment analysis (SSA) aims to automatically extract people’s opinions from a text in natural language and adequately represent that information in a graph structure. One of the most accurate methods for performing SSA was recently proposed and consists of approaching it as a dependency graph parsing task. Although we can find in the literature how transition-based algorithms excel in different dependency graph parsing tasks in terms of accuracy and efficiency, all proposed attempts to tackle SSA following that approach were based on graph-based models. In this article, we present the first transition-based method to address SSA as dependency graph parsing. Specifically, we design a transition system that processes the input text in a left-to-right pass, incrementally generating the graph structure containing all identified opinions. To effectively implement our final transition-based model, we resort to a Pointer Network architecture as a backbone. From an extensive evaluation, we demonstrate that our model offers the best performance to date in practically all cases among prior dependency-based methods, and surpasses recent task-specific techniques on the most challenging datasets. We additionally include an in-depth analysis and empirically prove that the average-case time complexity of our approach is quadratic in the sentence length, being more efficient than top-performing graph-based parsers.

This paper presents the first systematic benchmark evaluating Large Language Models (LLMs), specifically GPT-2, GPT-Neo-125M, and LLaMA-3.2-1B, as standalone classifiers for intrusion detection, covering both binary and multiclass classification tasks, using structured Zeek logs derived from the CIC IoT 2023 dataset. We compare their performance against established and widely used Machine Learning (XGBoost, Random Forest, Decision Tree) and Deep Learning models (MLP, GRU, LeNet-5) across key evaluation metrics: detection effectiveness (precision, recall and F1-score), inference speed, and resource consumption. All models are consistently trained and rigorously evaluated on the CIC IoT 2023 dataset, ensuring fair, reproducible, and transparent comparisons. Our findings indicate that while LLMs achieve strong F1-score exceeding 95%, and do not fully utilize available GPU resources, they still do not outperform top-performing ML models. Notably XGBoost achieves a higher F1-score of 96.96%, using only 4% of the available CPU. These results emphasize the practical trade-offs between detection capability, inference efficiency, and hardware requirements when applying LLMs in flow-based IDS contexts, particularly in resource-constrained environments such as IoT or edge deployments.

In recent years, research on the integration of evolutionary multi-objective optimization and (hyper-) heuristics (MOHHs) has significantly grown. This paper presents a comprehensive survey of MOHH research, categorizing existing approaches into four main classes: selection, generation, portfolio, and configuration MOHHs. Each category is analyzed in terms of methodology, key contributions, and open challenges. The analysis reveals an imbalance in research focus, with selection and portfolio MOHHs receiving the most attention, followed by configuration MOHHs, while generation MOHHs remain largely unaddressed. Selection MOHHs are further divided by the hierarchy of components they control: low-level approaches (which typically manage evolutionary operators) require further study on move acceptance methods, whereas mid-level approaches (which typically manage multi-objective evolutionary algorithms) need deeper exploration of selection strategies. Generation MOHHs, primarily based on genetic programming and grammatical evolution, lack investigation into alternative methodologies. Portfolio MOHHs, which produce a set of non-dominated constructive (hyper-) heuristics based on performance trade-offs, have been predominantly applied to combinatorial problems and exhibit limited diversity in the use of MOEAs as underlying optimizers. Configuration MOHHs, which focus on configuring algorithmic components for multi-objective optimizers, have largely relied on a single performance indicator, leaving room for multi-criteria performance approaches. Beyond this, the paper also reviews the test problems and practical applications that have been addressed by MOHHs, and outlines potential avenues for future research in the field.

Group fairness in machine learning is an important area of research focused on achieving equitable outcomes across different groups defined by sensitive attributes such as race or gender. Federated learning, a decentralized approach to training machine learning models across multiple clients, amplifies the need for fairness methodologies due to its inherent heterogeneous data distributions that can exacerbate biases. The intersection of federated learning and group fairness has attracted significant interest, with 48 research works specifically dedicated to addressing this issue. However, no comprehensive survey has specifically focused on group fairness in Federated Learning. In this work, we analyze the key challenges of this topic, propose practices for its identification and benchmarking, and create a novel taxonomy based on criteria such as data partitioning, location, and strategy. Furthermore, we analyze broader concerns, review how different approaches handle the complexities of various sensitive attributes, examine common datasets and applications, and discuss the ethical, legal, and policy implications of group fairness in FL. We conclude by highlighting key areas for future research, emphasizing the need for more methods to address the complexities of achieving group fairness in federated systems.

The rapid urbanization process has presented complex challenges that require innovative strategies to enhance urban living and promote sustainable growth. In this context, the concept of smart cities has quickly evolved, illustrating urban environments that utilize advanced technology to achieve greater efficiency, sustainability, and an improved quality of life for residents. The development of these smart environments relies on technologies like the Internet of Things (IoT), which collects extensive data through sensors, and Artificial Intelligence (AI), for advanced data processing and decision-making. For the latter, while traditional AI solutions have improved urban systems in multiple ways, emerging Generative Artificial Intelligence (GenAI) models signify a new era for smart cities, offering breakthroughs in urban design, simulation, and personalized, context-aware solutions. This article explores the applications, impacts, challenges, and promising future trends of GenAI within the context of smart cities, discussing generative urban intelligence perspectives for simulating alternative urban scenarios, co-designing infrastructure prototypes, and improving service delivery. It provides a pioneering perspective on an underexplored field that is expected to transform urban design, planning, and management.

The relentless growth of connected devices is transforming industrial, urban and domestic environments, yet it also expands the attack surface for distributed denial of service (DDoS), unauthorized access and data manipulation. Centralized security architectures struggle to cope with the scale and heterogeneity of the Internet of Things, creating single points of failure and privacy risks. This review takes a close look at how blockchain and artificial intelligence (AI) can work together to solve these problems. Blockchain plays an important role in decentralizing trust, maintaining data integrity, and enabling transparent audit trails. AI subfields such as machine learning (ML), deep learning (DL), reinforcement learning (RL), and multi-agent systems (MAS) enhance these benefits. They enable real-time anomaly detection, predictive analytics, and adaptive policy control. A seven axis Blockchain–AI Security Integration Schema (BASIS) is proposed to classify solutions by security objectives, intelligence modalities, trust primitives, deployment choices, scalability techniques, privacy controls and interoperability mechanisms. In this study also review Layer-2 consensus protocols, federated learning and lightweight deep learning models that address energy and computational constraints. Case studies from supply chains, healthcare and smart grids illustrate the benefits and limitations of current deployments. The evidence suggests that while AI improves the accuracy and responsiveness of threat detection, blockchain offers tamper-proof data provenance. However, there are still issues in achieving scalability, reducing computational overhead, and striking a balance between auditability and privacy. Hybrid on-chain/off-chain architectures, quantum-safe cryptography, and standardized frameworks to guarantee adoption and interoperability are some future research avenues.

Depth estimation methods for autonomous driving application face numerous challenges, such as capturing fine details and handling varying lighting conditions. Based on these challenges, LRDepth is proposed to improve the depth estimation task, which includes a simple high frequency enhancement module (HFEM) and a progressive residual denoising diffusion (PRDD) module. HFEM aids in extracting high-frequency components and amplifying the features, such as object edge details, generating more precise depth predictions. Inspired by the strong performance of diffusion models in various vision tasks, PRDD is designed to refine the depth predictions by reducing noise and enhancing edge details, which ensures the accurate representation of distant objects and subtle features. Extensive experiments on the KITTI and DIODE datasets demonstrated that the proposed network boosts the performance of monocular depth estimation, achieving more accurate long range depth predictions and improving model robustness in various lighting environments. The experiment results verified the method's adaptability, and the model is potential for real-world applications, which is beneficial for the optimization of visual perception module in intelligent driving system.

Mental health disorders are becoming a major global health concern and pose a significant burden on global healthcare systems. Nearly one billion people suffer from mental disorders, accounting for 13% of the global disease burden and $1 trillion in annual productivity loss. Depression is the leading cause of disability and suicide is the second leading cause of death among young individuals. Economic uncertainty, social isolation, climate change, shifting societal norms, political conflict, and increasing violence are key factors contributing to the high prevalence of mental health issues. In the future, increasing poverty and inequality are likely to worsen this trend, resulting in a greater incidence and burden of mental illness. Therefore, timely diagnosis and intervention are a high priority. Traditional diagnostic and intervention methods, such as self-report questionnaires, clinical interviews, psychotherapy, medication, electroconvulsive therapy, and occupational therapy, have drawbacks including subjectivity, time commitment, and the potential for prolonged treatment. Due to these limitations, advanced approaches are needed to improve diagnostic accuracy and precision and to develop more effective interventions. This review aims to explore and evaluate the applications of Artificial Intelligence in the diagnosis and treatment of mental health conditions. This study provides a thorough analysis of various artificial intelligence-driven techniques and their advancements in the diagnosis of mental health conditions. Artificial intelligence has the potential to greatly improve the accuracy and effectiveness of mental health conditions. Moreover, this work consolidates the research gaps in current techniques and provides research hypotheses on how to overcome the gaps using a proposed 3-tier solution.

This study explores the integration of directional changes (DC), genetic programming (GP), and multi-objective optimisation (MOO) to develop advanced algorithmic trading strategies. Directional changes offer a dynamic, event-based approach to market analysis, identifying significant price movements and trends. Genetic programming evolves trading rules to discover effective and profitable strategies. However, financial trading presents a multi-objective challenge, balancing conflicting objectives such as returns and risk. We propose a novel algorithmic trading framework, termed MOO3, which integrates genetic programming with the NSGA-II multi-objective optimisation algorithm to optimise three fitness functions: total return, expected rate of return, and risk. While the use of NSGA-II itself is well-established, our contribution lies in how we apply it within a trading context that combines (i) directional changes, (ii) genetic programming with both DC-based and physical-time indicators, and (iii) a modified Sharpe Ratio for post-optimisation strategy selection based on trader preferences. Utilising indicators from both paradigms allows the GP algorithm to create profitable trading strategies, while the multi-objective fitness function allows it to simultaneously optimise for risk. A definitive strategy is chosen from Pareto-optimal solutions using the modified Sharpe Ratio, allowing traders to prioritise multiple objectives. Our methodology is tested on 110 stock datasets from 10 international markets, aiming to demonstrate that the multi-objective framework can yield superior trading strategies with lower risk. Results indicate that the MOO3 algorithm consistently and significantly outperforms single-objective optimisation (SOO) methods, even when the same SOO criterion is employed for choosing a single, definitive investment strategy from the Pareto front.