Abstract: Reliable identification of maize leaf diseases is critical for mitigating crop losses, particularly in regions where farmers have limited access to experts. Although vision transformers (ViTs) have recently demonstrated strong performance in image recognition, their weak inductive bias and limited modelling of local texture patterns make them non-ideal for fine-grained maize leaf disease classification. To address these limitations, we propose ConvDeiT-Tiny, a lightweight hybrid ViT that improves DeiT-Ti by placing depthwise convolutions in parallel with multi-head self-attention modules in the first three transformer blocks. The local and global features captured by the convolution and attention modules are concatenated along the embedding dimension and fused using a multilayer perceptron. This results in richer token representations without significantly increasing model size. Across three datasets, ConvDeiT-Tiny (6.9M parameters) consistently outperformed DeiT-Ti, DeiT-Ti-Distilled, and DeiT-S (21.7M parameters) when trained from scratch. With transfer learning, ConvDeiT-Tiny achieved an accuracy of 99.15%, 99.35%, and 98.60% on the CD&S, primary, and Kaggle datasets, respectively, surpassing many previous studies with far fewer parameters. For explainability, we present gradient-weighted transformer attribution visualizations showing the disease lesions driving model predictions. These results indicate that injecting local inductive bias in early transformer blocks is beneficial for accurate maize leaf disease classification.

Abstract: This study addresses the highly dynamic runtime environment of microservice systems, the complex inter-service dependencies, and the hidden nature of anomalous behaviors. It proposes an anomaly detection method that integrates a meta learning mechanism. Based on multi-source monitoring data, the microservice execution process is modeled as a continuously evolving state sequence. A unified representation learning strategy is used to capture system evolution under normal conditions. The degree of state deviation is then adopted as the basis for anomaly discrimination. During modeling, different services or operating scenarios are treated as independent tasks. A meta learning framework is introduced to learn model initializations with strong transferability. This allows the model to adapt rapidly to new service instances and runtime environments under limited observations. It mitigates the impact of anomaly data scarcity and distribution shift. Compared with traditional methods that rely on fixed rules or single-scenario training, the proposed approach emphasizes shared runtime mechanism features. It maintains stable discrimination under noise interference and workload fluctuations. Comparative analysis under a unified data setting shows superior overall accuracy and more consistent anomaly discrimination compared with several representative methods. These results demonstrate strong robustness and generalization. The findings indicate that introducing meta learning into microservice anomaly detection improves adaptability and stability in complex cloud native environments. It provides an effective modeling strategy for anomaly identification in intelligent operations scenarios.

Abstract: Machine Learning (ML) is fundamentally reshaping education, offering tools to personalize instruction, automate assessment, and predict student outcomes. This paper provides a comprehensive overview of ML's role in education, tracing its evolution from early computer-assisted instruction to today's generative artificial intelligence (AI). We explore key applications, including intelligent tutoring systems, early warning systems for at-risk students, and automated essay scoring, highlighting their potential to address the long-standing challenge of individualized learning at scale. However, this technological integration is fraught with significant challenges. Ethical concerns regarding algorithmic bias, data privacy, and the "black box" nature of complex models threaten to exacerbate existing educational inequities. The recent proliferation of generative AI, exemplified by tools like ChatGPT, has further disrupted traditional paradigms of assessment and academic integrity, prompting urgent questions about the nature of learning itself. By synthesizing current research, this paper argues that while ML holds immense transformative promise, its successful and equitable implementation depends not on technological prowess alone, but on a concerted, ethically-grounded effort involving educators, researchers, and policymakers to ensure these tools augment human expertise and serve all learners.

Abstract:

We introduce the NeuroCore framework, a formal mathematical treatment of modular neural architectures in which a minimal executive Core—possessing no higher cognitive capabilities—autonomously orchestrates a heterogeneous collection of specialist modules through learned continuous-representation interfaces. The Core’s behavior is governed by two neuromodulation-inspired subsystems: a Dopamine System implementing distributional reinforcement learning with prediction-error intrinsic motivation and a stagnation penalty, and a Serotonin System formulated as a meta-reinforcement-learning controller that learns to optimize long-horizon constraint satisfaction. We make four theoretical contributions. First, we formalize the stagnation-modification tradeoff—proving that without explicit anti-stagnation pressure, optimal policies in self-modifying systems converge to modification-avoidance, and deriving the conditions under which the stagnation penalty restores non-trivial self-modification behavior (Theorem 1). Second, we prove a general non-convergence result for coupled self-modifying multi-objective systems, showing that the joint optimization does not admit guaranteed convergence to fixed points or bounded attractors in the parameter space (Theorem 2). Third, we establish partial stability guarantees: bounded representational drift via homeostatic Lyapunov functions (Theorem 3), local convergence under frozen modules via two-timescale stochastic approximation (Proposition 1), and modification frequency bounds (Proposition 2). Fourth, we derive information-theoretic costs for module manipulation operations that serve as principled proxies for true disruption. We propose seven falsifiable empirical predictions and discuss implications for the design of autonomous self-organizing AI systems.

Abstract: By decoupling services and enabling elastic deployment, microservice architecture improves system scalability and evolutionary capability. At the same time, it substantially increases operational complexity. Failures often exhibit cross service propagation and a mismatch between observed symptoms and underlying root causes. To address the heterogeneity and fragmentation of multi source observability data such as logs, metrics, and distributed traces, this study proposes a unified modeling and intelligent root cause localization method for microservice systems. The approach treats each service as a basic modeling unit and maps heterogeneous observations into a shared representation space. Service dependency structure is explicitly incorporated to characterize system state at a global level. Through structure aware modeling on the dependency graph, anomaly information is propagated and constrained along real invocation relations. This design enables more accurate separation of local disturbances from structural anomalies. In addition, a consistency based measure derived from state deviation is constructed to score service anomalies. Dependency relations are then used for attribution and ranking, which unifies root cause localization and impact analysis within a single framework. Comparative results show that the proposed method achieves more stable and consistent advantages across multiple evaluation metrics. It captures anomaly propagation patterns in microservice systems more effectively and provides a unified and structure aware solution for intelligent diagnosis of complex distributed systems.

Abstract: Interpreting and summarizing complex structured tabular data, particularly in specialized domains such as Korean administration, presents significant challenges due to intricate structures and domain-specific terminology. While Large Language Models (LLMs) offer promising capabilities, their direct application often results in information loss and misinterpretation. Existing solutions frequently necessitate extensive and resource-intensive model fine-tuning. To address these limitations, we propose Hierarchical Context-Aware Summarization (HCAS), a novel framework utilizing sophisticated prompt engineering and multi-stage reasoning. HCAS generates high-quality, human-friendly explanatory summaries for highlighted regions within complex Korean administrative tables, critically, without requiring large-scale model fine-tuning. It deconstructs the task into three distinct stages: Contextual Key Information Extraction, Explanatory Narrative Skeleton Construction, and Fluency and Readability Optimization, progressively enriching contextual understanding and refining output quality. Our comprehensive experiments on the NIKL Korean Table Explanation Benchmark demonstrate that HCAS consistently achieves superior performance, surpassing traditional fine-tuning methods and advanced in-context learning baselines on leading Korean LLMs. Further analyses validate HCAS's ability to produce factually accurate, coherent, and professionally appropriate summaries, while offering significant advantages in efficiency and resource utilization.

Abstract: Multi-horizon price forecasting is central to portfolio allocation, risk management, and algorithmic trading, yet deep learning architectures have advanced faster than rigorous financial benchmarking can properly evaluate them. Existing comparisons are often limited by inconsistent hyperparameter budgets, single-seed evaluation, narrow asset coverage, and a lack of statistical validation. This study presents a controlled comparison of nine architectures—Autoformer, DLinear, iTransformer, LSTM, ModernTCN, N-HiTS, PatchTST, TimesNet, and TimeXer—spanning four model families (Transformer, MLP, CNN, and RNN), evaluated across three asset classes (cryptocurrency, forex, and equity indices) and two forecasting horizons (h ∈ {4, 24} hours), for a total of 918 experiments. All runs follow a five-stage protocol: fixed-seed Bayesian hyperparameter optimization, configuration freezing per asset class, multi-seed final training, uncertainty-aware metric aggregation, and statistical validation. ModernTCN achieves the best mean rank (1.333) with a 75% first-place rate across 24 evaluation settings, followed by PatchTST (2.000), and the global leaderboard reveals a clear three-tier performance structure. Variance decomposition shows architecture explains 99.90% of raw RMSE variance versus 0.01% for seed randomness, and rankings remain stable across horizons despite 2–2.5× error amplification. Directional accuracy is statistically indistinguishable from 50% across all 54 model–category–horizon combinations, indicating that MSE-trained architectures lack directional skill at hourly resolution. These findings suggest that large-kernel temporal convolutions and patch-based Transformers consistently outperform alternatives, architectural inductive bias matters more than raw capacity, three-seed replication is sufficient, and directional forecasting requires explicit loss-function redesign; all code, data, trained models, and evaluation outputs are released for independent replication.

Abstract: Standard accuracy benchmarks evaluate whether a language model produces correct outputs but not whether it produces them consistently. We demonstrate that accuracy and output predictability are independent dimensions (Pearson r = -0.24, p = 0.56, N = 8 medical LLMs) when evaluated at a critical clinical summarization position. This independence yields a four-class behavioral taxonomy: IDEAL (convergent and accurate), EMPTY (convergent but inaccurate), DIVERGENT (high variance with incomplete outputs), and RICH (moderate variance with high accuracy).The DIVERGENT class exhibits stochastic incompleteness—summaries that are factually accurate but randomly incomplete across trials, with zero hallucinations. LAD occlusion, a critical clinical finding in STEMI cases, appears in only 22% of Llama 4 Scout summaries despite the model correctly identifying it when directly queried. This failure mode is invisible to standard benchmarks that average across outputs rather than measuring trial-to-trial variance.We propose a two-dimensional framework (Predictability × Accuracy) as a minimum requirement for clinical AI assessment, identify specific models unsuitable for deployment (Llama 4 Scout with Variance Ratio = 7.46; Llama 4 Maverick with Variance Ratio = 2.64), and flag one model requiring safety filter reconfiguration (Gemini Flash, 16% accuracy due to over-refusal). These findings demonstrate that current single-metric evaluation approaches systematically miss critical safety failures in clinical AI systems.

Abstract: Modern insurance organizations have adopted artificial intelligence in narrow, task-specific roles, resulting in fragmented systems that optimize isolated functions without fundamentally reshaping the underwriting and claims lifecycle. This “incrementalism” yields a human-default, sequential process plagued by structural bottlenecks, inconsistent risk evaluation, and limited transparency. This paper introduces NEXUS (Next-Generation Executive Underwriting and Settlement Intelligence), a framework to re-architect insurance as an AI-native system. NEXUS transitions AI from a peripheral tool to the primary orchestrator of end-to-end processes, conceptualizing the insurance lifecycle as a conversational, agent-orchestrated workflow. It is realized through a unified conversational interface that coordinates a decentralized ecosystem of specialized, collaborative AI agents each responsible for domain-specific reasoning such as geospatial risk assessment, financial verification, or medical outcome analysis. The central innovation is the Truth Score Engine (TSE), a governance-first aggregation mechanism that non-linearly synthesizes agent outputs by weighting evidentiary provenance, confidence estimates, and cross-agent consistency. The TSE governs decisions via a Three-Tiered Confidence Protocol: • High Confidence (>90%) validates outcomes for immediate human sign-off without re-verification; • Medium Confidence (60-90%) routes decision summaries for targeted human review of specific flags; • Low Confidence (<60%) escalates cases as ‘’Risky,’’ reverting to traditional manual investigation. This protocol yields a single, auditable decision artifact while preserving full traceability of the reasoning pathway. By embedding multi-agent coordination, contextual awareness, and tiered governance at the architectural level, NEXUS demonstrates a scalable pathway toward adaptive, transparent insurance systems. It ensures precision, combats fraud, and dramatically reduces settlement time, positioning AI-native governance as a foundational requirement for deploying trusted, autonomous decision-making in high-stakes financial domains.

Abstract: This paper presents the application of a micro unmanned aerial vehicle (UAV) that acts as a pollination agent in a controlled environment simulating greenhouse conditions. The micro-UAV system was integrated with a convolutional neural network (CNN) for autonomous flower detection and navigation. The ResNet-18 CNN architecture is was utilized onboard to perform real-time binary classification to accurately distinguish flowers from non-flower objects. The fusion of this deep-learning-based detection with precise micro-UAV navigation enables efficient identification and approaches to target flowers within optimal operational distances. Experimental evaluations revealed that the micro-UAV’s onboard camera combined with CNN processing outperformed standard webcams in terms of detection speed and accuracy, demonstrating the benefits of specialized hardware. Within the experiment, micro-UAV was pre-programmed to follow a ‘cross’-shaped flight pattern. Experimental result shows that the proposed system successfully detecting multiple flowers autonomously between distance of 30.5 cm to 91.5 cm within 149.1 seconds. Overall, this study validated the integration of neural network capabilities with micro-UAV navigation. These findings are crucial for drawing attention to the potential of neural network-enabled micro-UAVs as effective pollinators in enclosed agricultural environments and addressing the challenges faced by natural pollinators in greenhouses.

Abstract: Dyslexia impacts 5–15% of school-aged children globally, but automated screening mechanisms to detect it are rare, and such tools are relatively scarce in non-Latin scripts. The work introduces a bilingual deep learning model for dyslexia preliminary diagnosis through digitalized handwriting samples in both English and Arabic. Two computational methods were employed and compared systematically: the page-oriented classification strategy and the character-oriented classification method. For Arabic, an EnhancedCNN architecture is proposed to classify whole-page scans end-to-end by coping with cursive script and contextual letter forms. Both a baseline SimpleCNN model and a MobileNetV3-Small transfer learning model were trained on segmented letter crops from 123,554 labeled English samples. Preprocessing steps included the removal of instructor annotations, the Otsu adaptive thresholding method binarization and morphological processing noise removal and stroke refinement. Grad-CAM visualizations were included for model transparency and education decision aids, showing discriminative regions in page-level as well as character-level predictions. Experimental results proved that the proposed Arabic page-level model obtained 77% test accuracy, which constitutes preliminary proof of concept for AI-driven dyslexia screening in Arabic. English character-level approach using MobileNetV3 achieved 99% accuracy on the single letter detection task. This work also contributes to one of the earliest AI-assisted reading screening systems which is specifically designed for detecting dyslexia in Arabic script and brings systematic evidence on comparing hybrid page- and letter-level strategies for bilingual handwriting analysis.

Abstract: This study proposes a novel distributionally robust portfolio optimization framework based on Wasserstein generative modeling, aiming to address the challenges of distributional uncertainty, tail risk, and structural drift in financial markets. The model integrates Wasserstein distance-based robust optimization with generative adversarial learning to jointly enhance risk control and return stability. Specifically, a Wasserstein generative adversarial network is employed to reconstruct the latent distribution of asset returns, enabling the capture of non-Gaussian features and tail dependencies in complex market environments. By constructing an uncertainty set under the Wasserstein metric, the optimization process achieves dynamic balance between empirical risk minimization and robustness to distributional perturbations. Furthermore, the framework incorporates a dual optimization mechanism that alternately updates generative and optimization parameters to adaptively align with changing market structures. Experimental evaluations on multi-asset datasets demonstrate that the proposed model achieves higher Sharpe ratios, lower maximum drawdowns, and improved robustness compared with conventional reinforcement learning-based and mean-variance methods. The results verify that integrating Wasserstein generative modeling into distributionally robust optimization provides an effective and interpretable pathway for achieving stable asset allocation and risk-aware decision-making under volatile financial conditions.

Abstract: Social media platforms have established themselves as relevant sources of real-time information for urban traffic analysis. This study proposes an intelligent framework for the classification and spatiotemporal analysis of traffic incidents based on data flows constructed for controlled validation, based on real reports from platforms such as X and Telegram. The approach integrates adaptive machine learning and incremental density-based clustering. An Adaptive Random Forest (ARF) incremental classifier is used to identify the type of incident, allowing for continuous updating of the model in response to changes in traffic flow and concept drift. The classified events are then processed using DenStream, a clustering algorithm that incorporates a temporal decay mechanism designed to identify dynamic spatial patterns and discard older information. The evaluation is performed in a controlled streaming simulation environment that replicates the dynamics of cities such as Panama and Guayaquil, using prequential evaluation metrics. The results suggest that this hybrid architecture is a viable approach for urban traffic monitoring, providing useful information for Intelligent Transportation Systems (ITS) by processing authentic social signals.

Abstract: Cybercriminals have increasingly leveraged sophisticated techniques to bypass traditional signature-based detection systems through the use of Ransomware-as-a-Service (RaaS) platforms, double and triple extortion strategies, and advanced evasion mechanisms. As a result, ransomware attacks have reached unprecedented levels. Using this systematic evaluation framework, we examine the current state and effectiveness of machine learning (ML) and deep learning (DL) approaches for ransomware detection, addressing critical gaps in existing research methodologies while providing comprehensive recommendations for future research. The study analyses multiple AI paradigms including supervised learning algorithms such as Random Forests and Support Vector Machines, unsupervised techniques such as clustering and anomaly detection, and deep learning architectures such as Convolutional Neural Networks and Long Short-Term Memory networks. Hybrid approaches combining static and dynamic analysis consistently achieve superior performance, with accuracy rates exceeding 99% when properly implemented. As part of the framework, fundamental challenges are addressed such as dataset quality and diversity, feature extraction and selection methodologies, data preprocessing techniques, and performance evaluation metrics that have been tailored specifically for cybersecurity applications. Several findings indicate that ensemble learning methods outperform individual classifiers, with Random Forest algorithms being particularly effective at handling high-dimensional feature spaces while maintaining interpretability for security analysts. As a result of the study, significant limitations have been identified in current research, including an overreliance on static data sets that do not capture evolving threat landscapes, an inadequate representation of modern attack vectors, and a limited ability to generalize across different operational environments. Future directions of this research include explainable AI integration for transparent decision-making, adaptive real-time detection systems, and federated learning approaches for collaborative threat intelligence sharing while maintaining organizational privacy. It provides standardized methodologies for data curation, feature engineering, model development, and performance benchmarking, enabling fair comparisons between different AI approaches and facilitating reproducible research. This work contributes to essential guidance for cybersecurity practitioners, policymakers, and researchers in developing robust, adaptive, and interpretable ransomware detection systems capable of defending against increasingly sophisticated cyber threats while considering ethical concerns and regulatory compliance requirements in modern digital ecosystems.

Abstract: In the era of connected and cyber-physical systems, firmware and software releases management in industries like IoT, smart grids, medical devices, UAVs, satellites, and vehicles has been growing more complex. These systems require constant, secure, and reliable updates to keep up with increasing user expectations and regulatory requirements. However, the growing complexity of release pipelines, as well as increasing security risks and tight regulatory demands, make it a challenging process. The emergence of Generative AI (GenAI), and Large Language Models (LLMs) in particular, provides a transformative opportunity for innovating traditional release management processes and provides the ability to automate documentation, triaging defects, creating risk assessment, and compliance.This article is aimed at discussing a GenAI-augmented vision for release management in critical systems that provides a conceptual framework built on current literature in OTAs updates, cyber security, and AI in product development. Through a conceptual and integrative review, we provide a number of use case scenarios, in which GenAI can be advantageously used to strengthen release pipelines, such as automation for root cause analysis (RCA) reports, intelligent defect triage, and risk mitigation strategies. We also address the ethical considerations and reliability risks associated with integrating AI in safety-critical environments, including emphasizing the importance of human oversight and governance.

Abstract: Assessing skill levels from videos of human activities is critical for applications in sports coaching, surgical training, and workplace safety. Existing approaches typically assign a global skill score to a video, failing to localize where and how skilled performers differ from novices. We propose SkillDiff, a framework that quantifies fine-grained skill differences between paired demonstration videos at the temporal segment level. Our method first aligns expert and novice videos temporally through a learned alignment module, then computes per-segment skill difference embeddings that capture deviations in execution quality, timing efficiency, and motion patterns. SkillDiff introduces: (1) a Temporal Alignment Backbone that establishes dense frame correspondences between demonstrations of varying skill, (2) a Differential Skill Encoder that transforms alignment residuals into interpretable skill difference features, and (3) a Segment-Level Scoring Head that produces localized quality assessments. Experiments on BEST, Fis-V, and AQA-7 benchmarks show that SkillDiff achieves state-of-the-art correlation with expert annotations (Spearman rho=0.93 on BEST), while providing temporally localized feedback that existing global scoring methods cannot.

Abstract: Since 2011, the mandatory adoption of Korean International Financial Reporting Standards (K-IFRS) by listed Korean firms has improved the consistency of financial reporting and enhanced comparability across firms and over time This institutional change has made it more feasible to construct long-horizon firm–year panel datasets and apply quantitative predictive analyses. The KoTaP dataset provides standardized firm–year panel data for Korean listed non-financial firms over 2011–2024, and this study empirically evaluates the feasibility of risk screening based on one-year-ahead (t→t+1) forecasting of tax-avoidance proxies (CETR, GETR, TSTA, TSDA) using KoTaP. Specifically, we define an ex-ante setting in which only information observable in year t is used to predict tax-avoidance indicators at t+1. We then propose a leakage-free evaluation protocol that enforces chronological splits and fits all preprocessing steps on the training data only. We further partition input features into raw and derived variables and compare three configurations Raw-only, Derived-only, and Raw+Derived to quantify the contribution of derived feature construction. Finally, we compare three machine-learning models and one deep-learning model under the same evaluation procedure and derive practical implications for model selection and deployment in terms of performance and stability.

Abstract: This study focuses on the tendency of agents in long-horizon sequential tasks to rely on short-term states and to underutilize historical information, and proposes a cognitive modeling and learning framework with long-term memory and reasoning capabilities. The framework provides a unified cognitive description of the agent's decision process. It introduces a structured long-term memory mechanism to support continuous storage and selective updating of cross-temporal key information. On this basis, a memory retrieval-driven reasoning module is constructed so that experience can explicitly participate in the formation of current decision logic. To address the separation between memory and decision making in conventional policy models, the framework tightly couples perception representation, memory management, reasoning processes, and policy generation into an end-to-end cognitive loop. This design strengthens goal consistency and behavioral stability in long-horizon interactive environments. Comparative evaluations in open source interactive task settings demonstrate consistent advantages in task completion quality, decision efficiency, and long-term information utilization. The results indicate that the proposed cognitive modeling framework effectively mitigates decision difficulties caused by long-range dependencies and partial observability. Overall, the study shows that integrating long-term memory and reasoning within a unified learning framework is an important approach for improving sustained decision-making capability in complex environments.

Abstract: The rapid growth of course-based online platforms has expanded access to learning in non-formal educational contexts. However, many existing systems primarily emphasize content delivery and transactional functionality, with limited pedagogical coordination between activity tracking, assessment, and learner guidance. This paper proposes a multi-layered architectural framework for orchestrating course-based online learning through a structured separation of content management, learning activity tracking, competency modeling, adaptive support, and credentialing. A key contribution of the proposed model is the distinction between content completion and competency mastery. While completion indicators reflect engagement, mastery estimation is grounded in assessment evidence aligned with formally defined learning outcomes. The architecture introduces an orchestration mechanism realized through distributed coordination across layers rather than through a centralized autonomous controller, positioning artificial intelligence as a supportive tool for assessment item generation and feedback provision. A functional web-based prototype, SoftLearner, has been developed to validate the feasibility of the infrastructure and learning layers. The higher layers are defined architecturally and provide a foundation for future empirical validation. The proposed framework contributes to the development of pedagogically oriented AI-supported systems in non-formal online learning environments and supports cumulative, competency-based recognition of learning outcomes within lifelong learning contexts.

Abstract: Forecasting stock prices, market trends, and associated emotions remains a complex challenge due to the market's inherent volatility, nonlinearity, susceptibility to factors such as news events, and the limited availability of financial data. Stock prices are noisy, unpredictable, and sensitive to investor sentiment \cite{ko2021, dahal2023, kumarsh2025}. To account for this, this study considers three models that effectively capture both the spatial relationships (similar sector stocks) between stocks and the temporal trends in their price movements, thereby addressing limitations in traditional forecasting methods. Specifically, we consider semiconductor stocks, which are known for their high volatility and strong correlation to technological advancements. The proposed models include Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks, a GNN-based architecture, and a Spatio-Temporal Graph Neural Network (ST-GNN). These models are trained, tested, and validated on datasets that combine historical data with sentiment insights from financial news sources. It utilizes a GRU-based temporal module to improve recognition of evolving market patterns. We demonstrate our model's ability to adapt to dynamic financial environments and predict whether a stock closes at a higher or lower price than at the trading day open.