Partial multi-label learning (PML) aims to learn from samples annotated with incomplete label sets, where only a subset of the ground-truth labels is available. This problem becomes particularly challenging in resource-constrained environments, where model efficiency and interpretability are essential. In this paper, we propose a novel framework, Partial Multi-label Green Learning with Global Attention Disambiguation (PMGL-GAD), which integrates lightweight unsupervised feature extraction with global attention-guided disambiguation. Specifically, we adopt the green learning architecture PixelHop++ to construct multi-scale hierarchical features in a feed-forward, training-free manner. To enhance discriminative capability for partial label resolution, we introduce a Global Attention Disambiguation (GAD) module into the green learning pipeline. The GAD module builds a cross-channel data subspace via global attention mechanisms and a learnable gating strategy, enabling feature reweighting across labels without deep network training. Furthermore, a disambiguated label subspace is learned concurrently to improve label assignment. Experimental results demonstrate that PMGL-GAD achieves competitive performance with significantly reduced complexity, offering a promising direction for interpretable, low-carbon multi-label learning in weakly supervised scenarios.
This study aims to provide a comprehensive and structured overview of how DL frameworks have been utilized for the recognition and classification of white blood cell (WBC) images. The review is guided by three primary research questions: (1) What are the dominant trends in DL model adoption for WBC classification? (2) How do image preprocessing and segmentation techniques influence performance? (3) What are the current limitations and prospects in this domain? A systematic literature review was conducted following the Preferred Reporting Items for Systematic Reviews and Meta-Analyses guidelines. Studies were selected and analyzed based on relevance, methodological rigor, and contribution to the field. An analytical framework comprising 18 technical dimensions was applied to evaluate the DL pipelines used. Convolutional neural networks (CNN) were found to dominate the field, used in 81.6
The rapid expansion of multivariate time series (MTS) data has made deep learning a central tool for forecasting across industrial and scientific domains. As these architectures move into settings such as clinical decision support, industrial monitoring, and financial risk management, researchers need to examine not only predictive accuracy but also computational cost and the kinds of explanations that a model can support. This survey re-examines deep learning models for MTS forecasting through the requirements of efficiency and explainability. Our contributions are threefold: (1) We synthesize the literature through a dual-axis scalability view over sequence length L and variate dimension N, together with a three-question explainability taxonomy covering temporal importance, variate importance, and pattern decomposition. (2) We review dominant architectural families, including Transformer variants, frequency-domain methods, linear and channel-independent backbones, state space models, and foundation-style forecasters, to describe how different designs occupy different efficiency–explainability profiles under specific data and deployment assumptions. (3) Building on this synthesis, we identify key open challenges—including the absence of standardized explainability benchmarks for time series, the interpretability gap in state space models, and the need to advance from correlational to causal explanations—and offer reporting considerations to guide more comparable future research.
Effective Traffic Signal Control (TSC) in large-scale transportation networks is essential for enhancing urban mobility, reducing congestion, and improving safety. However, traditional control methods often fail to effectively address the complexity, dynamic conditions, and multimodal demands of modern urban traffic systems. In recent years, Reinforcement Learning (RL) has emerged as a promising solution for achieving adaptive and scalable TSC. This paper presents a systematic and up-to-date review of RL-based methods for large-scale TSC. We analyze representative studies published between 2013 and 2025, presenting a comprehensive analysis of traffic simulation environments, transportation modalities, and advances in methodologies. Key aspects include multi-agent paradigms, state and action representations, reward mechanisms, RL frameworks, as well as advanced representation learning and cooperative strategies for large-scale transportation networks. We also provide a critical discussion on performance evaluation and opportunities for improvement, and conclude by summarizing the current challenges and outlining future research directions. This review aims to inform and guide the development of next-generation RL-based TSC systems that promote sustainable, safe, and efficient urban transportation.
Reinforcement Learning (RL) is a foundational framework in Artificial Intelligence (AI) that enables agents to acquire optimal decision-making strategies through interactions with their environments. Building on principles of trial-and-error learning, RL adapts dynamically by leveraging feedback in the form of rewards or penalties. This paper provides a comprehensive survey of RL, its integration with Deep Learning into Deep Reinforcement Learning (DRL), and the emerging field of model-based planning. It systematically examines a wide range of algorithms, from classical tabular methods to state-of-the-art DRL techniques, while evaluating their scalability, sample efficiency, and applicability in various domains. The survey offers an in-depth analysis of the strengths and limitations of these approaches, along with practical guidance for selecting and implementing RL algorithms. Key challenges, such as convergence, stability, and the exploration-exploitation trade-off, are thoroughly addressed to tackle real-world constraints. Additionally, this work explores the transformative impact of RL across diverse fields, including robotics, autonomous systems, intelligent transportation, energy management, and financial modeling. By bridging theoretical advancements and practical applications, this survey aims to provide researchers and practitioners with a comprehensive understanding of RL’s capabilities for solving complex and dynamic problems in AI and beyond.
Training machine learning models with more than one data modality has enhanced predictive performance in most contexts. Thus, many recent applications of machine learning use data from different sources and forms. Multimodal data augmentation (MMDA) addresses critical challenges in multimodal learning, such as data scarcity, modality imbalance, and cross-modal alignment. This survey systematically reviews 68 state-of-the-art MMDA approaches, and, as result, proposes a taxonomy for the area. For each revised paper, this survey analyzes its methodology, applications, and predictive performance gains, while highlighting key challenges, such as scalability and evaluation metrics. The proposed taxonomy provides a unified framework for understanding MMDA methods, their strengths, and limitations. This survey also identifies emerging trends, including the integration of large language models and diffusion processes, and outlines future research directions to advance multimodal learning.
Recent advances in artificial intelligence, exemplified by AlphaGo-like systems, have demonstrated the effectiveness of integrating neural networks with classical tree search for complex decision-making. By serving as learned heuristics, neural networks provide informative priors to guide the search process, while search-based planning improves decision quality through structured exploration and long-horizon optimization, with the resulting high-quality solutions further distilled into neural models to form a bidirectional decision-making framework. Moreover, the integration of learned environment models has further expanded the applicability of this paradigm beyond board games to diverse domains, such as de novo drug design and large language model inference. In this survey, we provide a systematic review of neural-guided heuristic tree search from the perspective of deep bidirectional decision-making. Representative classical tree search algorithms and their integration with neural heuristics are revisited, and recent advances are synthesized in terms of three fundamental challenges: optimality degradation under neural guidance, reduced search efficiency due to imperfect heuristics, and the difficulty of learning accurate and search-compatible heuristic functions. Moreover, representative real-world applications are reviewed to demonstrate the practical applicability of neural-guided tree search, followed by a discussion of open challenges and promising directions for future research. By synthesizing these developments within a unified framework, this survey clarifies the methodological foundations and empirical advances of neural-guided heuristic tree search, while highlighting its potential for general-purpose decision-making.
Real-time object detection is a cornerstone of intelligent transportation systems, where the YOLO (You Only Look Once) family has long defined the accuracy–speed trade-off. The emergence of Mamba—a selective state-space model that captures long-range dependencies in linear time—has triggered a wave of research embedding selective state spaces into YOLO detectors, promising global context without the quadratic cost of self-attention. Yet existing visual-Mamba surveys remain narrative and treat detection as one application among many; none isolates the YOLO-specific design space or assesses it under a reproducible protocol. Following the PRISMA-2020 methodology, this review screens 1053 Scopus records published between December 2023 and May 2026 and synthesizes 114 high-quality studies, 14 of them recovered through a documented two-pass re-evaluation procedure. We organize the field along a structured taxonomy of variant family, integration position, and application domain. The synthesis reveals pronounced terminology fragmentation—three-quarters of studies introduce custom Mamba variants—identifies YOLOv8, v11, and v5 as the dominant base frameworks, and shows that integration clusters at the feature-fusion, backbone, and attention positions. Critically, a Pareto analysis of 39 studies demonstrates that benchmark heterogeneity invalidates direct cross-study ranking, while medical imaging stays underexplored and only 33
Financial time-series forecasting lies between AI and market microstructure, but most studies optimise generic error metrics instead of risk-adjusted economic value under realistic frictions. Unlike NLP and vision, the field lacks a shared, reviewer-enforced standard for data handling and evaluation, leading to persistent problems such as data leakage, backtest overfitting and metric-chasing on RMSE/MAE. This paper introduces QFRS a novel, enforceable by reviewers and editors, seven-standard framework and checklist for evaluating and reporting financial asset forecasting and trading claims. QFRS covers quantitative studies on equities (stocks), forex, cryptocurrencies, rates, derivatives (futures, forwards, options, swaps), energy prices, and commodities (gold, oil and silver) and other asset classes. The seven standards specify an end-to-end experimental pipeline, covering (i) dataset construction, (ii) labelling, (iii) point-in-time feature engineering, (iv) leakage-free scaling or normalisation, (v) time-respecting data splits, (vi) evaluation metrics and (vii) cost and slippage-aware backtesting with explicit execution assumptions and decision rules mapping predictions to positions. To validate the standard’s diagnostic value, a compliance audit of Scopus-indexed forex forecasting papers published in 2025 is presented. None of these papers achieved full compliance across all seven standards, with economic backtesting (12.2
Large language models (LLMs) are increasingly being explored in geoscience, where scientific knowledge is expressed through specialized terminology, heterogeneous documents, maps, imagery, geospatial structures, and temporally ordered interpretations. This review critically examines the emerging literature on geoscience-oriented LLMs (GeoLLMs), focusing on the tasks, construction strategies, evaluation needs, and unresolved challenges that distinguish them from generic LLM applications. We first synthesize the GeoLLM task landscape, including geological information extraction and semantic normalization, relation modeling and knowledge graph construction, evidence-grounded question answering, multimodal map–image–text reasoning, and high-value applications such as hazard-related information analysis, mineral prospectivity evidence synthesis, and chronostratigraphic interpretation. We then review model construction and adaptation strategies, including geoscience corpus engineering, parameter-efficient tuning, domain-adaptive pretraining, retrieval augmentation, ontology and knowledge-graph grounding, and multimodal representation learning. Across these studies, a consistent theme is that GeoLLM outputs should be assessed not only by linguistic fluency, but also by terminology consistency, evidence traceability, spatial and temporal coherence, multimodal grounding, uncertainty expression, and expert validation. The review further identifies major open challenges, including fragmented data and benchmarks, regional and multilingual terminology variation, scale-aware multimodal reasoning, causal and spatiotemporal consistency, hallucination, overtrust, data privacy, proprietary-model dependence, and deployment governance. By using geoscience as a demanding application domain rather than a universal testbed, this review clarifies where domain-specific LLMs can add value, where current evidence remains limited, and what evaluation and governance practices are needed for reliable scientific use.
As the integration of Artificial Intelligence into network intrusion detection systems matures, a critical gap remains in the rigorous empirical benchmarking of datasets, preprocessing techniques, and model effectiveness. This article presents a comparative experimental study of anomaly and threat detection techniques used in network analysis through a multistep pipeline. First, we perform a structured comparison and Exploratory Data Analysis of the most commonly used network security datasets to quantitatively assess their balance, diversity, and real-world representativeness. Subsequently, we experimentally evaluate the performance of distinct Machine Learning, Deep Learning, and Hybrid Models derived under standardized preprocessing techniques to determine the most effective combinations for specific attack vectors. Our results demonstrate that hybrid architectures achieve superior generalisation, yet face challenges regarding computational overhead and cross-dataset adaptability. Addressing these limitations, we propose a proof-of-concept adaptive architecture designed to handle concept drift and adversarial threats. Finally, we outline a roadmap for future research, emphasizing the necessity of dynamic, verifiable AI-driven systems that operate reliably in evolving cybersecurity environments.
This paper undertakes a systematic investigation of the medical image segmentation benchmark datasets, which play a crucial role in the notable progress of medical image segmentation task. The datasets serve as the foundational infrastructure comparable to a backbone that supports and drives the development of medical image segmentation. Consequently, examination of these datasets emerges as a critical topic in research. In order to address the current lack of a systematic summary and thorough analysis of benchmark datasets for medical image segmentation, and to gain insights into their current status and future trends, this survey consolidates and categorizes the fundamental aspects of these benchmark datasets from five perspectives: (1) organ/tissue segmentation datasets; (2) lesion region segmentation datasets; (3) multi-organ/tissue segmentation datasets; (4) multi-lesion region segmentation datasets; (5) multi-organ/tissue and lesion region segmentation datasets. The survey elucidates the prevailing challenges and identifies potential avenues for future investigation. Additionally, a comprehensive review of the existing available dataset resources is also provided, including statistics from 140 datasets, covering seven core anatomical regions and over 60 subcategories of tissues/lesion types. Data from five segmentation task scenarios is incorporated into the dataset statistics. We aim to map out the full panorama of medical image segmentation datasets, serving as a comprehensive reference for researchers in this field and contributing to future studies. Related resources are available at: https://github.com/CVLife/Awesome-medical-segmentation-datasets.
Despite rapid digitalization, agricultural computer vision still faces persistent data bottlenecks. The collection and annotation of field images are constrained by seasonality and biological variability, require domain expertise, must adapt to the perception and navigation conditions of ground robots, and are further limited by privacy and data-sharing concerns. Together, these factors restrict the scale, diversity, and transferability of real-world datasets. This paper provides a systematic review of two-dimensional (2D) and three-dimensional (3D) synthetic data generation (SDG) pipelines in agriculture. We outline 2D pipelines ranging from classical augmentation and compositing to learning-based generative models, and we review 3D approaches spanning procedural plant modeling, photorealistic rendering engines, customized simulation platforms, and reconstruction-based assets. This review focuses on 2D and 3D SDG for agricultural vision tasks (image acquisition, object detection, semantic and instance segmentation, and classification), including weed and disease management, phenotyping, yield management, precision livestock farming, aquaculture, and fruit management. Beyond cataloguing methods, we adopt a deployment-oriented perspective and present task-driven design rules: select SDG according to the primary bottleneck (class imbalance, occlusion, or domain shift); favor a "3D geometry first, 2D stylization later" workflow when geometric fidelity and visual realism must both be preserved; combine domain randomization or translation with limited target-domain fine-tuning; and align method choice with available expertise and computing resources. We also synthesize challenges related to realism and diversity, generalization and sim-to-real transfer, and ethical and legal considerations. For evaluation, we compile common practices, including expert plausibility checks, quantitative image metrics, and controlled comparisons on identical real-world test sets, complemented by ablation studies and statistical tests. Ultimately, the review integrates methodological, empirical, and deployment perspectives into a practical roadmap for selecting, building, and validating SDG pipelines that remain agronomically credible and transferable to real agricultural environments.
Deep learning has become a key enabling technology for detecting security-relevant events in visual surveillance data acquired from CCTV systems, UAV platforms, and other imaging sensors. However, despite substantial progress in benchmark performance, the operational deployment of such systems remains challenging due to dataset bias, domain shift, limited robustness, edge-computing constraints, and a lack of operationally meaningful evaluation metrics. This structured narrative review synthesises research published between 2021 and 2026, complemented by selected seminal studies. It is guided by predefined research questions and a documented, purposive search and selection strategy, and examines deep learning approaches for weapon detection, violence detection, anomaly recognition, person-related security tasks, perimeter monitoring, and UAV-based surveillance. The analysis covers major architectural paradigms, including YOLO-based detectors, RT-DETR, Vision Transformers, spatiotemporal networks, anomaly-detection frameworks, and edge-optimised models, with particular attention to dataset characteristics, model robustness, adversarial vulnerabilities, multimodal sensing, model compression, and regulatory considerations associated with the EU AI Act. As structuring contributions, the review proposes a multi-dimensional taxonomy that links security-event categories to architecture class, evaluation metric, sensor modality, and EU AI Act risk level, together with a critical analysis of architecture-specific failure modes under operational conditions. It further identifies recurring limitations that hinder real-world deployment: across representative studies, evaluation remains dominated by accuracy-oriented metrics such as mAP, F1-score, and FPS, whereas operational aspects including detection latency, false-alarm burden, and deployment robustness are insufficiently addressed. To bridge this gap, the review proposes Time-to-Detection (TTD) as a complementary operational evaluation framework, and highlights persistent challenges related to realistic datasets, demographic and domain biases, adversarial resilience, privacy-preserving learning, and federated deployment. The findings indicate that future research should prioritise standardised operational benchmarking, TTD-aware evaluation, robust and explainable models, realistic security datasets, efficient edge-AI deployment, and regulation-aware system design. Addressing these challenges will be essential for translating advances in deep learning into reliable and trustworthy security applications.
In the twenty-first century, urbanization has become one of the most transformative processes, driving significant changes in land use, infrastructure, and environmental conditions. These dynamics underscore the need for accurate urban planning, environmental monitoring, and disaster management. Traditional approaches utilize optical and multispectral remote sensing imagery and often fail to distinguish spectrally similar urban materials. In contrast, hyperspectral imaging improves classification accuracy by providing far more detailed spectral and spatial information. This survey presents a comprehensive and systematic review of hyperspectral image change detection, with a particular focus on the emerging role of vision transformers (ViTs) and transformer-driven architectures. This study analyzes the evolution from traditional pixel-based and machine learning methods to deep learning paradigms, highlighting the limitations of convolutional neural networks (CNNs) in capturing long-range dependencies. The paper provides an in-depth discussion of HSI principles, urban data acquisition constraints, transformer architectures, spatial spectral attention mechanisms, hybrid CNN + ViT models, domain adaptation strategies, and interpretability techniques. Furthermore, we analyze recent advances in spatial-only, spectral-only, and hybrid attention mechanisms, CNN + ViT fusion models, and spatial-spectral tokenization strategies tailored for hyperspectral data. We further summarize benchmark hyperspectral datasets, multimodal data fusion with LiDAR, SAR, and multispectral imagery, synthetic dataset generation using generative adversarial networks, interpretability and explainability techniques, and the computational challenges of cloud and edge deployment. The reviewed deep learning models are organized into six methodological families, i.e., CNN-based, transformer-based, graph-based, domain-adaptive and label-efficient, fusion-based, and lightweight designs, and a unified experimental evaluation compares representative models under a common protocol. By synthesizing experimental trends and comparative performance across extensive benchmark studies, this survey highlights current limitations, open challenges, and future research directions.
Mainstream AI research emphasises capability growth and tolerates low failure rates when average-case performance is high. AI safety and alignment research has a different mission: to ensure that catastrophic failures never occur, under sparse evidence, adversarial dynamics, and fat-tailed risk. We argue that the two domains differ along two analytically independent axes — capability profile (demonstrating the absence of hazardous behaviours versus the presence of positive capabilities) and risk profile (bounding worst-case outcomes under fat-tailed uncertainty versus optimising average-case performance) — and that mainstream epistemic practices are inadequate on both. Building on a structured synthesis grounded in a preregistered bibliometric baseline, we identify five cross-cutting gap dimensions in current alignment research, including the near-absence of institutionalised independent verification. To address these gaps we propose ECAISA, an Epistemic Code for AI Safety and Alignment comprising eight principles, a three-level scoring rubric, a four-level disclosure ladder that reconciles transparency with information-hazard and commercial-confidentiality constraints, a tiered applicability scheme, an infohazard adjudication procedure, and seven anti-gaming mechanisms. ECAISA does not certify that any AI system is safe; it constrains how safety-relevant research claims are documented, checked, and relied upon, with auditability rather than certification as its governance target. A retrospective rubric audit (κ = 0.79) demonstrates instrument feasibility; a four-stage validation roadmap is proposed.
Multivariate time series (MTS) analysis is increasingly important for extracting insights from complex, interdependent temporal data in domains such as healthcare, finance, and industrial monitoring. Recent advances in deep learning have significantly improved MTS modeling; yet, the rapidly expanding literature remains fragmented across tasks, architectures, and evaluation practices. This survey concentrates on deep learning-centric approaches in MTS research across seven key tasks: classification, clustering, forecasting, anomaly detection, imputation, representation learning, and transfer learning. To organize the field, we introduce a multi-dimensional taxonomy that categorizes methods along three complementary axes: (i) modeling and learning paradigms (e.g., Transformers, self-supervised contrastive learning, generative models, and emerging LLM-based approaches), (ii) downstream analytical objectives, and (iii) deployment-oriented capabilities such as scalability, robustness to irregular sampling and distribution shifts, interpretability, and transferability. Beyond conceptual synthesis, we conduct a targeted reproducibility audit by re-evaluating representative models using public codebases under controlled experimental conditions. Our findings show that while broad performance trends are often reproducible, exact numerical results can vary across models, datasets, and protocols, particularly when preprocessing or seed-level details are underspecified. Lastly, we identify key challenges, including high dimensionality, sparsity, non-stationarity, and concept drift, and outline seven research directions toward more scalable, reliable, and interpretable MTS systems.
Plant disease detection and severity estimation are crucial to sustainable agricultural productivity and global food security, necessitating the need for efficient and accurate diagnostic tools. This paper systematically analyzes 137 studies using the PRISMA 2020 framework, focusing on deep learning methods used in detecting and estimating plant disease severity. The review covers classification, detection, segmentation, and regression approaches to diagnosing plant diseases and quantifying severity, detailing the transition in architecture design from traditional CNN to hybrid CNN-transformer frameworks. The findings show that the best precision in severity estimation is achieved by using a segmentation approach, while regression strikes a good balance between accuracy and performance, whereas classification stands out as the most scalable model. However, there is still a performance drop from laboratory to the field settings due to the problems of domain shift, class imbalance, and long-tail distributions. This paper highlights domain adaptation, few-shot learning, and multimodal fusion as the key techniques to achieve better generalizability. Robotic implementation and explainable AI are also discussed as crucial aspects in implementing practical disease diagnostic systems. This work provides a system-level synthesis to guide the development of robust, interpretable, and field-ready plant disease diagnostic systems.
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.
The rapid advancement of deep learning models (DLMs) in computer vision is driven by their capacity to unify feature representation and task inference into end-to-end automated frameworks, thereby surpassing manual feature engineering. This survey categorizes DLMs into Convolutional Neural Networks (CNNs), Visual Transformers (ViTs), and hybrid models based on their core components (convolutional vs. attention-based), systematically reviews their technical evolution, and critically analyzes their strengths and limitations. CNNs, the pioneering paradigm, leverage the locality of convolution kernels for efficient hierarchical pattern extraction. However, their reliance on multi-layer stacking to aggregate global context and rigidity to geometric variations has motivated innovations such as deformable convolutions to expand receptive fields. ViTs, emerging later, prioritize the globality of attention mechanisms to model long-range dependencies directly, but face computational bottlenecks and noise sensitivity, which are addressed by sparse attention and hierarchical tokenization. Hybrid models now dominate as optimal compromises, combining the locality of CNNs and globality of ViTs to balance efficiency and expressiveness. Moreover, we emphasize universal techniques—neural architecture search (NAS), model compression, and self-supervised learning—as indispensable tools to enhance robustness and adaptability across architectures. Despite progress, challenges persist in interpretability, data efficiency, and real-world robustness. We advocate prioritizing lightweight hybrid designs, explainable attention-convolution combinations, and cross-modal systems integrating more data beyond vision. This survey not only systematically reviews the technical evolution of DLMs but also proposes a forward-looking roadmap for adaptive, scalable, and trustworthy vision systems, grounded in current technological trajectories.
