AI, Volume 6, Issue 5 (May 2025) – 19 articles

Neural Ordinary Differential Equations (Neural ODEs) have emerged as a promising approach for learning the continuous-time behaviour of dynamical systems from data. However, Neural ODEs are black-box models, posing challenges in interpreting and understanding their decision-making processes. This raises concerns about their application in critical domains such as healthcare and autonomous systems. To address this challenge and provide insight into the decision-making process of Neural ODEs, we introduce the eXplainable Neural ODE (XNODE) framework, a suite of eXplainable Artificial Intelligence (XAI) techniques specifically designed for Neural ODEs. Drawing inspiration from classical visualisation methods for differential equations, including time series, state space, and vector field plots, XNODE aims to offer intuitive insights into model behaviour. Although relatively simple, these techniques are intended to furnish researchers with a deeper understanding of the underlying mathematical tools, thereby serving as a practical guide for interpreting results obtained with Neural ODEs. The effectiveness of XNODE is verified through case studies involving a Resistor–Capacitor (RC) circuit, the Lotka–Volterra predator-prey dynamics, and a chemical reaction. The proposed XNODE suite offers a more nuanced perspective for cases where low Mean Squared Error values are obtained, which initially suggests successful learning of the data dynamics. This reveals that a low training error does not necessarily equate to comprehensive understanding or accurate modelling of the underlying data dynamics. Full article

This study addresses the problem of authorship attribution with a novel method for modeling writing style using dependency tree subtree parsing. This method exploits the syntactic information of sentences using mixed syntactic n-grams (mixed sn-grams). The method comprises an algorithm to generate mixed sn-grams by integrating words, POS tags, and dependency relation tags. The mixed sn-grams are used as style markers to feed Machine Learning methods such as a SVM. A comparative analysis was performed to evaluate the performance of the proposed mixed sn-grams method against homogeneous sn-grams with the PAN-CLEF 2012 and CCAT50 datasets. Experiments with PAN 2012 showed the potential of mixed sn-grams to model a writing style by outperforming homogeneous sn-grams. On the other hand, experiments with CCAT50 showed that training with mixed sn-grams improves accuracy over homogeneous sn-grams, with the POS-Word category showing the best result. The study’s results suggest that mixed sn-grams constitute effective stylistic markers for building a reliable writing style model, which machine learning algorithms can learn. Full article

Background/Objectives: The resilience of safety-critical systems is gaining importance due to the rise in cyber and physical threats, especially within critical infrastructure. Traditional static resilience metrics may not capture dynamic system states, leading to inaccurate assessments and ineffective responses to cyber threats. This work aims to develop a data-driven, adaptive method for resilience metric learning. Methods: We propose a data-driven approach using inverse reinforcement learning (IRL) to learn a single, adaptive resilience metric. The method infers a reward function from expert control actions. Unlike previous approaches using static weights or fuzzy logic, this work applies adversarial inverse reinforcement learning (AIRL), training a generator and discriminator in parallel to learn the reward structure and derive an optimal policy. Results: The proposed approach is evaluated on multiple scenarios: optimal communication network rerouting, power distribution network reconfiguration, and cyber–physical restoration of critical loads using the IEEE 123-bus system. Conclusions: The adaptive, learned resilience metric enables faster critical load restoration in comparison to conventional RL approaches. Full article

In this article, we propose a robust star classification methodology leveraging light curves collected from 15 datasets within the Kepler field in the visible optical spectrum. By employing a Bagging neural network ensemble approach, specifically an Bagging-Performance Approach Neural Network (BAPANN), which integrates three supervised neural network architectures, we successfully classified 760 samples of curves which represent 9 type of stars. Our method demonstrated a high classification accuracy of up to 97% using light curve datasets containing 13, 20, 50, 150, and 450 points per star. The BAPANN achieved a minimum error rate of 0.1559 and exhibited efficient learning, requiring an average of 29 epochs. Additionally, nine types of stellar variability were classified through 45 conducted tests, taking into account error margins of 0, 5, and 10 for the light curve samples. These results highlight the BAPANN model’s robustness against uncertainty and ability to converge quickly in terms of iterations needed for learning, training, and validation. Full article

In today’s high-stakes arenas—from healthcare to defense—algorithms are advancing at an unprecedented pace, yet they still lack a crucial element of human decision-making: an instinctive caution that helps prevent harm. Inspired by both the protective reflexes seen in military robotics and the human amygdala’s role in threat detection, we introduce a novel idea: an integrated module that acts as an internal “caution system”. This module does not experience emotion in the human sense; rather, it serves as an embedded safeguard that continuously assesses uncertainty and triggers protective measures whenever potential dangers arise. Our proposed framework combines several established techniques. It uses Bayesian methods to continuously estimate the likelihood of adverse outcomes, applies reinforcement learning strategies with penalties for choices that might lead to harmful results, and incorporates layers of human oversight to review decisions when needed. The result is a system that mirrors the prudence and measured judgment of experienced clinicians—hesitating and recalibrating its actions when the data are ambiguous, much like a doctor would rely on both intuition and expertise to prevent errors. We call on computer scientists, healthcare professionals, and policymakers to collaborate in refining and testing this approach. Through joint research, pilot projects, and robust regulatory guidelines, we aim to ensure that advanced computational systems can combine speed and precision with an inherent predisposition toward protecting human life. Ultimately, by embedding this cautionary module, the framework is expected to significantly reduce AI-induced risks and enhance patient safety and trust in medical AI systems. It seems inevitable for future superintelligent AI systems in medicine to possess emotion-like processes. Full article

The rapid advancement of artificial intelligence (AI) and digital transformation is reshaping labor markets, emphasizing creativity as a core competency in entrepreneurship education. Large Language Models (LLMs) provide personalized learning experiences through natural language processing (NLP), enhancing real-time feedback and problem-solving skills. However, research on how LLMs foster creativity in entrepreneurship education remains limited. This study analyzed the technical characteristics and educational impact of LLMs, focusing on their applications in entrepreneurship education and their role in fostering creativity-driven learning environments. Specifically, it explores the educational effects of LLMs, their integration into entrepreneurship education, and the ways in which they enhance learners’ creative thinking. A systematic literature review using the PRISMA methodology was conducted to analyze existing studies. Findings suggest that LLMs improve self-efficacy, cognitive engagement, and creative problem-solving, supporting entrepreneurship education in areas such as business model development, market analysis, and multicultural communication. Despite these benefits, concerns remain regarding over-reliance, ethical risks, and the need for critical thinking frameworks. This study proposes a hybrid model integrating LLMs with traditional pedagogies to maximize creativity. Future research should explore long-term effects, cross-cultural applications, and ethical challenges to ensure responsible implementation. Full article

Background: The increasing prevalence of cyber threats in industrial cyber–physical systems (ICPSs) necessitates advanced solutions for threat detection and analysis. This research proposes a novel GPT-based framework for assessing cyber threats, leveraging artificial intelligence to process and analyze large-scale cyber event data. Methods: The framework integrates multiple components, including data ingestion, preprocessing, feature extraction, and analysis modules such as knowledge graph construction, clustering, and anomaly detection. It utilizes a hybrid methodology combining spectral residual transformation and Convolutional Neural Networks (CNNs) to identify anomalies in time-series cyber event data, alongside regression models for evaluating the significant factors associated with cyber events. Results: The system was evaluated using 9018 cyber-related events sourced from 44 global news portals. Performance metrics, including precision (0.999), recall (0.998), and F1-score (0.998), demonstrate the framework’s efficacy in accurately classifying and categorizing cyber events. Notably, anomaly detection identified six significant deviations during the monitored timeframe, starting from 25 September 2023 to 25 November 2024, with a sensitivity of 75%, revealing critical insights into unusual activity patterns. The fully deployed automated model also identified 11 correlated factors and five unique clusters associated with high-rated cyber incidents. Conclusions: This approach provides actionable intelligence for stakeholders by offering real-time monitoring, anomaly detection, and knowledge graph-based insights into cyber threats. The outcomes highlight the system’s potential to enhance ICPS security, supporting proactive threat management and resilience in increasingly complex industrial environments. Full article

Occupancy detection for large buildings enables optimised control of indoor systems based on occupant presence, reducing the energy costs of heating and cooling. Through machine learning models, occupancy detection is achieved with an accuracy of over 95%. However, to achieve this, large amounts of data are collected with little consideration of which of the collected data are most useful to the task. This paper demonstrates methods to identify if data may be removed from the imbalanced time-series training datasets to optimise the training process and model performance. It also describes how the calculation of the class density of a dataset may be used to identify if a dataset is applicable for data reduction, and how dataset fusion may be used to combine occupancy datasets. The results show that over 50% of a training dataset may be removed from imbalanced datasets while maintaining performance, reducing training time and energy cost by over 40%. This indicates that a data-centric approach to developing artificial intelligence applications is as important as selecting the best model. Full article

Background: Accurate ground segmentation in 3D point clouds is critical for robotic perception, enabling robust navigation, object detection, and environmental mapping. However, existing methods struggle with over-segmentation, under-segmentation, and computational inefficiency, particularly in dynamic or complex environments. Methods: This study proposes FASTSeg3D, a novel two-stage algorithm for real-time ground filtering. First, Range Elevation Estimation (REE) organizes point clouds efficiently while filtering outliers. Second, adaptive Window-Based Model Fitting (WBMF) addresses over-segmentation by dynamically adjusting to local geometric features. The method was rigorously evaluated in four challenging scenarios: large objects (vehicles), pedestrians, small debris/vegetation, and rainy conditions across day/night cycles. Results: FASTSeg3D achieved state-of-the-art performance, with a mean error of <7%, error sensitivity < 10%, and IoU scores > 90% in all scenarios except extreme cases (rainy/night small-object conditions). It maintained a processing speed 10× faster than comparable methods, enabling real-time operation. The algorithm also outperformed benchmarks in F1 score (avg. 94.2%) and kappa coefficient (avg. 0.91), demonstrating superior robustness. Conclusions: FASTSeg3D addresses critical limitations in ground segmentation by balancing speed and accuracy, making it ideal for real-time robotic applications in diverse environments. Its computational efficiency and adaptability to edge cases represent a significant advancement for autonomous systems. Full article

Currently, large language models (LLMs) have been utilized in many aspects of natural language processing. However, due to their significant size and high computational demands, large computational resources are required for deployment. In this research, we focus on the automated approach for size reduction of such a model. We propose the framework to perform the automated pruning based on combinatorial optimization. Two techniques were particularly studied, i.e., particle swarm optimization (PSO) and whale optimization algorithm (WOA). The model pruning problem was modeled as a combinatorial optimization task whose the goal is to minimize model size while maintaining model accuracy. The framework systematically explores the search space to identify the most optimal pruning configurations, removing redundant or non-contributory parameters. The two optimizations, PSO and WOA, were evaluated for their ability to efficiently navigate the search space. As a result, with PSO, the proposed framework can reduce the model size of Llama-3.1-70B by 13.44% while keeping the loss of model accuracy at 19.25%; with WOA, the model size reduction is 12.07% with 22.81% loss of model accuracy. Since accuracy degradation may occur during pruning process, the framework integrates the post-process to recover the model accuracy. After this process, the pruned model loss can reduce to 12.72% and 14.83% using PSO and WOA, respectively. Full article

Background: Investment decisions in stocks are one of the most complex tasks due to the uncertainty of which stocks will increase or decrease in their values. A diversified portfolio statistically reduces the risk; however, stock choice still substantially influences the profitability. Methods: This work proposes a methodology to automate investment decision recommendations with clear explanations. It utilizes generative AI, guided by prompt engineering, to interpret price predictions derived from neural networks. The methodology also includes the Artificial Intelligence Trust, Risk, and Security Management (AI TRiSM) model to provide robust security recommendations for the system. The proposed system provides long-term investment recommendations based on the financial fundamentals of companies, such as the price-to-earnings ratio (PER) and the net margin of profits over the total revenue. The proposed explainable artificial intelligence (XAI) system uses DeepSeek for describing recommendations and suggested companies, as well as several charts based on Shapley additive explanation (SHAP) values and local-interpretable model-agnostic explanations (LIMEs) for showing feature importance. Results: In the experiments, we compared the profitability of the proposed portfolios, ranging from 8 to 28 stock values, with the maximum expected price increases for 4 years in the NASDAQ-100 and S&P-500, where both bull and bear markets were, respectively, considered before and after the custom duties increases in international trade by the USA in April 2025. The proposed system achieved an average profitability of 56.62% while considering 120 different portfolio recommendations. Conclusions: A t-Student test confirmed that the difference in profitability compared to the index was statistically significant. A user study revealed that the participants agreed that the portfolio explanations were useful for trusting the system, with an average score of 6.14 in a 7-point Likert scale. Full article

Background. Single-cell RNA sequencing (scRNA-seq) has transformed genomics by enabling the study of cellular heterogeneity. However, its high dimensionality, noise, and sparsity pose significant challenges for data analysis. Methods. We investigate the use of Hyperdimensional Computing (HDC), a brain-inspired computational framework recognized for its noise robustness and hardware efficiency, to tackle the challenges in scRNA-seq data analysis. We apply HDC to both supervised classification and unsupervised clustering tasks. Results. Our experiments demonstrate that HDC consistently outperforms established methods such as XGBoost, Seurat reference mapping, and scANVI in terms of noise tolerance and scalability. HDC achieves superior accuracy in classification tasks and maintains robust clustering performance across varying noise levels. Conclusions. These results highlight HDC as a promising framework for accurate and efficient single-cell data analysis. Its potential extends to other high-dimensional biological datasets including proteomics, epigenomics, and transcriptomics, with implications for advancing bioinformatics and personalized medicine. Full article

Generating character-consistent and personalized dialogue for Non-Player Characters (NPCs) in Role-Playing Games (RPGs) poses significant challenges, especially due to limited memory retention and inconsistent character representation. This paper proposes a framework for generating personalized dialogues based on character-specific knowledge. By combining static knowledge fine-tuning and dynamic knowledge graph technology, the framework generates dialogue content that is more aligned with character settings and is highly personalized. Specifically, the paper introduces a protective static knowledge fine-tuning approach to ensure that the language model does not generate content beyond the character’s cognitive scope during conversations. Additionally, dynamic knowledge graphs are employed to store and update the interaction history between NPCs and players, forming unique “experience-response” patterns. During dialogue generation, the paper first parses player input into an Abstract Meaning Representation (AMR) graph, retrieves relevant memory nodes from the knowledge graph, and constructs a fused graph structure. This integrated graph is encoded via a graph neural network to generate high-dimensional semantic vectors, which are then used to retrieve and supplement knowledge from the vector database. Ultimately, the model generates personalized responses consistent with the NPC’s identity. Experimental results demonstrate that the framework significantly enhances the authenticity of NPC dialogues and player immersion and performs well on multiple large-scale language models. Full article

Due to range of factors in the development stage, generative artificial intelligence (AI) models cannot be completely free from bias. Some biases are introduced by the quality of training data, and developer influence during both design and training of the large language models (LLMs), while others are introduced in the text-to-image (T2I) visualization programs. The bias and initialization at the interface between LLMs and T2I applications has not been examined to date. This study analyzes 770 images of librarians and curators generated by DALL-E from ChatGPT-4o prompts to investigate the source of gender, ethnicity, and age biases in these visualizations. Comparing prompts generated by ChatGPT-4o with DALL-E’s visual interpretations, the research demonstrates that DALL-E primarily introduces biases when ChatGPT-4o provides non-specific prompts. This highlights the potential for generative AI to perpetuate and amplify harmful stereotypes related to gender, age, and ethnicity in professional roles. Full article

Effective anomaly detection is essential for realizing modern and secure industrial control systems. However, the direct integration of anomaly detection within such a system is complex due to the wide variety of hardware used, different communication protocols, and given industrial requirements. Many components of an industrial control system allow direct integration, while others are designed as closed systems or do not have the required performance. At the same time, the effective usage of available resources and the sustainable use of energy are more important than ever for modern industry. Therefore, in this paper, we present a modular and hybrid concept that enables the integration of efficient and effective anomaly detection while optimising the use of available resources under consideration of industrial requirements. Because of the modular and hybrid properties, many functionalities can be outsourced to the respective devices, and at the same time, additional hardware can be integrated where required. The resulting flexibility allows the seamless integration of complete anomaly detection into existing and legacy systems without the need for expensive centralised or cloud-based solutions. Through a detailed evaluation within an industrial unit, we demonstrate the performance and versatility of our concept. Full article

(1) Background: In recent years, Transformer-based models have dominated the time-series forecasting domain, overshadowing recurrent neural networks (RNNs) such as Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU). While Transformers demonstrate superior performance, their high computational cost limits their practical application in resource-constrained settings. (2) Methods: In this paper, we reconsider RNNs—specifically the GRU architecture—as an efficient alternative to time-series forecasting by leveraging this architecture’s sequential representation capability to capture cross-channel dependencies effectively. Our model also utilizes a feed-forward layer right after the GRU module to represent temporal dependencies, and aggregates it with the GRU layers to predict future values of a given time-series. (3) Results and conclusions: Our extensive experiments conducted on different real-world datasets show that our inverted GRU (iGRU) model achieves promising results in terms of error metrics and memory efficiency, challenging or surpassing state-of-the-art models on various benchmarks. Full article

Background/Objectives: The implementation of artificial intelligence-based systems for disease detection using biomedical signals is challenging due to the limited availability of training data. This paper deals with the generation of synthetic EEG signals using deep learning-based models, to be used in future research for training Parkinson’s disease detection systems. Methods: Linear models, such as AR, MA, and ARMA, are often inadequate due to the inherent non-linearity of time series. To overcome this drawback, long short-term memory (LSTM) networks are proposed to learn long-term dependencies in non-linear EEG time series and subsequently generate synthetic signals to enhance the training of detection systems. To learn the forward and backward time dependencies in the EEG signals, a Bidirectional LSTM model has been implemented. The LSTM model was trained on the UC San Diego Resting State EEG Dataset, which includes samples from two groups: individuals with Parkinson’s disease and a healthy control group. Results: To determine the optimal number of cells in the model, we evaluated the mean squared error (MSE) and cross-correlation between the original and synthetic signals. This method was also applied to select the length of the hidden state vector. The number of hidden cells was set to 14, and the length of the hidden state vector for each cell was fixed at 4. Increasing these values did not improve MSE or cross-correlation and unnecessarily increased computational complexity. The proposed model’s performance was evaluated using the mean-squared error (MSE), Pearson’s correlation coefficient, and the power spectra of the synthetic and original signals, demonstrating the suitability of the proposed method for this application. Conclusions: The proposed model was compared to Autoregressive Moving Average (ARMA) models, demonstrating superior performance. This confirms that deep learning-based models, such as LSTM, are strong alternatives to statistical models like ARMA for handling non-linear, multifrequency, and non-stationary signals. Full article

(1) Background: The misuse of transformation technology using medical images is a critical problem that can endanger patients’ lives, and detecting manipulation via a deep learning model is essential to address issues of manipulated medical images that may arise in the healthcare field. (2) Methods: The dataset was divided into a real fundus dataset and a manipulated dataset. The fundus image manipulation detection model uses a deep learning model based on a Convolution Neural Network (CNN) structure that applies a concatenate operation for fast computation speed and reduced loss of input image weights. (3) Results: For real data, the model achieved an average sensitivity of 0.98, precision of 1.00, F1-score of 0.99, and AUC of 0.988. For manipulated data, the model recorded sensitivity of 1.00, precision of 0.84, F1-score of 0.92, and AUC of 0.988. Comparatively, five ophthalmologists achieved lower average scores on manipulated data: sensitivity of 0.71, precision of 0.61, F1-score of 0.65, and AUC of 0.822. (4) Conclusions: This study presents the possibility of addressing and preventing problems caused by manipulated medical images in the healthcare field. The proposed approach for detecting manipulated fundus images through a deep learning model demonstrates higher performance than that of ophthalmologists, making it an effective method. Full article

Advances in continuous glucose monitoring (CGM) technologies and wearable devices are enabling the enhancement of automated insulin delivery systems (AIDs) towards fully automated closed-loop systems, aiming to achieve secure, personalized, and optimal blood glucose concentration (BGC) management for individuals with diabetes. While model predictive control provides a flexible framework for developing AIDs control algorithms, models that capture inter- and intra-patient variability and perturbation uncertainty are needed for accurate and effective regulation of BGC. Advances in artificial intelligence present new opportunities for developing data-driven, fully closed-loop AIDs. Among them, deep reinforcement learning (DRL) has attracted much attention due to its potential resistance to perturbations. To this end, this paper conducts a literature review on DRL-based BGC control algorithms for AIDs. First, this paper systematically analyzes the benefits of utilizing DRL algorithms in AIDs. Then, a comprehensive review of various DRL techniques and extensions that have been proposed to address challenges arising from their integration with AIDs, including considerations related to low sample availability, personalization, and security are discussed. Additionally, the paper provides an application-oriented investigation of DRL-based AIDs control algorithms, emphasizing significant challenges in practical implementations. Finally, the paper discusses solutions to relevant BGC control problems, outlines prospects for practical applications, and suggests future research directions. Full article