Artificial Intelligence Review最新文献

This study investigates and enhances the Forward–Forward (FF) algorithm, a novel neural network training approach recently introduced by Geoffrey Hinton. Unlike Back-Propagation (BP), which relies on separate forward and backward passes, FF performs two forward passes, one with positive (real) data and one with negative data, potentially generated by the network itself. We evaluate FF across six diverse datasets to assess its general performance and employ two well-established and four state-of-the-art metaheuristic algorithms on four datasets to optimize its hyper-parameters. While FF offers a notable advantage in memory efficiency, making it suitable for low-resource hardware, it initially lags behind BP in predictive accuracy. Through automated hyper-parameter optimization, FF demonstrates substantial improvements, with validation accuracy increasing from 93% to 97% on Pneumonia-MedMnist, 87% to 91% on Fashion-MNIST, 82% to 94% on OrganC-MedMnist, and 88% to 95% on OrganA-MedMnist, while maintaining or reducing training time. These results demonstrate that metaheuristic optimization not only accelerates convergence but also significantly enhances FF’s predictive performance, bridging the gap between memory efficiency and accuracy.

Unsupervised state-learning aims to infer latent variables or representations from multivariate time series that can serve as predictive summaries of a system evolution. Depending on the modelling assumptions, these summaries may appear as discrete regimes, posterior belief states, continuous latent trajectories or learnt embeddings that are subsequently interpreted or regularised as states. State-learning enables segmentation, anomaly detection, predictive maintenance and clinical monitoring, but the literature is fragmented across probabilistic regime-switching models, clustering and change-point methods, as well as modern representation learning. This survey unifies these strands and provides a synthesis aimed at practical method selection. We formalise the notion of state and introduce a taxonomy that makes modelling choices explicit along key dimensions: discrete versus continuous states, observable versus latent representations, output formats (hard assignments, posteriors, embeddings) and fixed versus adaptive state cardinality. We then review three major methodological families: probabilistic models, clustering/segmentation approaches, as well as self-supervised or generative representation learning. We highlight core assumptions, computational trade-offs and failure modes under missingness, irregular sampling, non-stationarity, as well as scarce evaluation labels. Building on this synthesis, we present a method-selection framework and implementation-oriented workflows that summarise recurring patterns in the literature and our practical reading of them, covering preprocessing, model fitting, diagnostics and uncertainty handling. Furthermore, we propose an evaluation and reporting protocol tailored to state-learning, emphasising leakage-free temporal splits, temporal coherence, interpretability, complexity control and uncertainty calibration. Finally, we summarise widely used benchmark testbeds that provide evaluation-only labels.

Machine Learning (ML) has been widely applied across numerous domains due to its ability to automatically identify informative patterns from data for various tasks. The availability of large-scale data and advanced computational power enables the development of sophisticated models and training strategies, leading to state-of-the-art performance, but it also introduces substantial challenges. Quantum Computing (QC), which exploits quantum mechanisms for computation, has attracted growing attention and significant global investment as it may address these challenges. Consequently, Quantum Machine Learning (QML), the integration of these two fields, has received increasing interest, with a notable rise in related studies in recent years. We are motivated to review these existing contributions regarding quantum circuit-based learning models for classical data analysis and highlight the identified potentials and challenges of this technique. Specifically, we focus not only on QML models, both kernel-based and neural network-based, but also on recent explorations of their integration with classical machine learning layers within hybrid frameworks. Moreover, we examine both theoretical analysis and empirical findings to better understand their capabilities, and we also discuss the efforts on noise-resilient and hardware-efficient QML that could enhance its practicality under current hardware limitations. In addition, we cover several emerging paradigms for advanced quantum circuit design and highlight the adaptability of QML across representative application domains. This study aims to provide an overview of the contributions made to bridge quantum computing and machine learning, offering insights and guidance to support its future development and pave the way for broader adoption in the coming years.

Metacognition has gained significant attention for its potential to enhance autonomy and adaptability of artificial agents but remains a fragmented field: diverse theories, terminologies, and design choices have led to disjointed developments and limited comparability across systems. Existing overviews remain at a conceptual level that is undiscerning to the underlying algorithms, representations, and their respective effectiveness. We address this gap by performing a systematic scoping review. Reports were included if they described techniques enabling Computational Metacognitive Architectures (CMAs) to model, store, remember, and process their episodic metacognitive experiences, one of Flavell’s (1979) three foundational components of metacognition. Searches were conducted in 16 databases between December 2023 and June 2024. Data were charted using a 20-item framework considering pertinent aspects and analysed via structured narrative synthesis. A total of 101 reports on 35 distinct CMAs were included. Our findings show that metacognitive experiences may boost system performance and explainability, e.g., via self-repair. However, lack of standardisation and limited evaluations may hinder progress: only 17% of CMAs were quantitatively evaluated regarding this review’s focus, and significant terminological inconsistency limits cross-architecture synthesis. Systems also varied widely in memory content, data types, and employed algorithms. Limitations include the non-iterative nature of the search query, heterogeneous data availability, and an under-representation of sub-symbolic CMAs. Future research should focus on standardisation and evaluation, e.g., via community-driven challenges, and on transferring promising principles to emergent systems.

The advent of Artificial General Intelligence (AGI) promises to revolutionize affective computing. However, current research lacks a systematic framework that accounts for the inherently cross-space nature of human emotion, where physiological states (physical space), cultural norms (social space), cognitive processes (thinking space), and digital data (cyberspace) continuously interact and shape each other. To address this gap, this paper starts by repositioning affective computing within the theory of Cyber-Physical-Social-Thinking (CPST) space, conceptualizing it as a dynamic process of cross-space transformation rather than an isolated function. AGI capabilities of perception, learning, reasoning, memory, generalization, and generation serve as the computational engines that enable these cross-space transformations. We then systematically map AGI techniques (e.g., causal modeling, chain-of-thought, diffusion models) to specific affective computing tasks, showing how each technique supports their realization. By integrating CPST as a theoretical backbone and AGI as an enabling technology, this paper provides a systematic roadmap for next-generation affective computing. Potential applications and future directions are also discussed, highlighting how this framework advances human-level affective intelligence toward genuine empathy and contextual awareness.

Over the past years, embedding learning on networks has shown tremendous results in link prediction tasks for complex systems, with a wide range of real-life applications. Learning a representation for each node in a knowledge graph allows us to capture topological and semantic information, which can be processed in downstream analyses later. In the link prediction task, high-dimensional network information is encoded into low-dimensional vectors, which are then fed to a predictor to infer new connections between nodes in the network. As the network complexity (that is, the numbers of connections and types of interactions) grows, embedding learning turns out increasingly challenging. This review covers published models on embedding learning on multiplex networks for link prediction. First, we propose refined taxonomies to classify and compare 38 models (in 36 papers) across five methodological categories and three embedding classes. Second, we review and address the problem of reproducible and fair evaluation of embedding learning on multiplex networks for the link prediction task. Finally, we highlight challenges specific to performance evaluation on directed multiplex networks, and address them by introducing a novel and fair testing procedure. This review constitutes a crucial step towards the development of more performant and tractable embedding learning approaches for multiplex networks and their fair evaluation for the link prediction task. We also suggest guidelines on the evaluation of models, and provide an informed perspective on the challenges and tools currently available to address downstream analyses applied to multiplex networks.

Deploying Large Language Models at scale has shifted the dominant bottleneck from compute during training to memory and I/O during inference. As parameter counts reach hundreds of billions and context windows stretch past a million tokens, latency and throughput are limited not by arithmetic but by data movement across the memory hierarchy. This survey decomposes inference I/O into three flows—model weight I/O, Key-Value (KV) cache I/O, and activation I/O—and uses roofline analysis to map each optimization to the memory-hierarchy level it targets. We cover quantization, PagedAttention, FlashAttention, speculative decoding, KV cache compression, and offloading, alongside system-level orchestration (continuous batching, disaggregated prefill-decode, prefix caching) and hardware trends (HBM scaling, CXL, processing-in-memory, unified memory). A composability analysis reveals that stacking optimizations causes the dominant bottleneck to oscillate between weight and KV cache I/O. We close by identifying open problems in unbounded-context scaling, expert caching, edge deployment, and I/O-aware benchmarking.

Semantic role labeling (SRL) is a central natural language processing task for understanding predicate-argument structures within texts and enabling downstream applications. Despite extensive research, comprehensive surveys that critically synthesize the field from a unified perspective remain lacking. This survey makes several contributions beyond organizing existing work. We propose a unified taxonomy organized around four analytical perspectives that categorizes SRL research along model architectures, syntax feature modeling, application scenarios, and multimodal extensions. We provide a literature-based synthesis of when and why syntactic features help, identifying conditions under which syntax-aided approaches provide consistent gains over syntax-free counterparts. We offer a critical assessment of SRL in the era of large language models, examining the complementary roles of LLMs and specialized SRL systems and identifying directions for hybrid approaches. We extend the scope of SRL surveys to cover multimodal settings including visual, video, and speech modalities, and review structural differences in evaluation across these modalities. Methodologically, this work is a structured narrative survey rather than a formal systematic review: we describe explicit search procedures and inclusion criteria below, but do not follow PRISMA-style reporting protocols such as inter-reviewer screening or formal quality assessment of included studies. Literature was collected through structured searches of the ACL Anthology, IEEE Xplore, the ACM Digital Library, and Google Scholar, covering publications from 2000 to 2025, with the initial search completed in March 2025 and additional publications incorporated during the peer-review process through December 2025. Explicit inclusion and exclusion criteria were applied to yield approximately 200 primary references. SRL benchmarks, evaluation metrics, and paradigm modeling approaches are discussed alongside practical applications across domains. Future research directions are analyzed, addressing the evolving role of SRL with large language models and broader NLP impact.

Graph-based learning and explainable artificial intelligence (XAI) are increasingly used to improve both predictive performance and transparency in financial risk modelling. This paper presents a systematic literature review of AI and machine learning approaches for credit risk assessment and fraud detection, with specific attention to graph-based methods and explainable frameworks. Following a PRISMA-guided methodology, 149 studies published between 2015 and 2025 were analysed across multiple academic databases. The review identifies three key findings. First, graph-based models, particularly graph neural networks, can improve the modelling of relational dependencies in financial data, although their use remains more developed in fraud detection than in credit risk assessment. Second, XAI techniques such as SHAP, LIME, and rule-based methods are increasingly used to support interpretability, auditability, and regulatory compliance, but their integration with graph-based models remains limited. Third, recent research is shifting from purely predictive modelling towards deployment-oriented financial AI systems that address class imbalance, concept drift, scalability, real-time monitoring, and governance. To address these gaps, this paper proposes a deployment-oriented reference framework that links graph construction, graph-based modelling, explainability mechanisms, monitoring, human oversight, and governance controls. The findings provide a structured synthesis of current research and practical guidance for developing scalable, transparent, and accountable financial AI systems.