There is a significant interest in Artificial Intelligence (AI) systems for automating decision making. While there are several AI technology solutions to choose from, there is a need to understand the fundamentals of human decision making and conceptualizing required AI system architecture. However, the challenge is: how can we conceptualize the design of an AI system architecture? This paper proposes trimodal thinking for architecting human-centric AI systems. This paper reviews dual system thinking approach, fast and slow, from the well-known book, published by Daniel Kahneman, with a specific purpose to conceptualize the human-centric AI system architecture. This review found that the fast intuitive system 1 and slow self-controlled deliberate system 2 are useful for conceptualizing the design of the AI system architecture. However, there is a need for system 3 to adjust, balance, connect, regulate and switch the control between these two systems. System 3 is critical for mitigating system 1 unconscious and system 2 conscious biases, power imbalance and noise. Relevant concepts, from principle-agent theory, homeostasis, control theory, and the atomic human, were also used to guide the proposal of trimodal thinking. Thus, this paper proposed trimodal thinking and demonstrated its use for architecting human-centric AI systems using the combination of (1) fast, (2) slow and (3) control systems. This paper also highlighted several implications for broader enterprise architecture discipline and indicated new topics across decision science and computer science areas for further research. It is anticipated that this work will provide a necessary foundation for conceptualizing technology agnostic and human-centric AI systems architecture. This will support the complex AI system explainability and transparency needs of non-technical users and technical AI system engineers.

This article provides a research viewpoint on the training procedure of a Bidirectional Encoder Representations from Transformers (BERT) model starting from the beginning, employing the Tensor Processing Unit (TPU), with a particular emphasis on the Spanish language. This article focuses on the techniques and strategies, such as the whole word masking (WWM) technique, that are essential for achieving successful BERT training in multiple languages, including English. The goal is to provide efficient development of natural language processing applications for understanding the Spanish language. The BEThiz model exhibits exceptional proficiency in comprehending contextual cues and producing accurate responses for masked inputs, surpassing other models within the same category and language. This characteristic sets our model apart and solidifies its role as a particularly potent tool in the field of natural language processing.

To support the development of wearable medical devices for remote monitoring and treatment of cardiovascular diseases, we tackle the data scarcity problem that hinders the application of machine learning methods. We propose a self-supervised approach applied to cardiological signals, which benefits from existing datasets despite differences between them and inconsistencies within them. We develop a specific implementation: a diffusion autoencoder with a semantic encoder based on linear recurrent units, trained on ECG signals (various leads mixed together) without any annotations. The semantic encoder is evaluated as a feature extractor by measuring classification metrics of a logistic regression on a dataset not included in the self-supervised training. We obtain promising results and propose future directions.

Condition-based maintenance(CBM) has attracted widespread attention as it can help to reduce maintenance cost and enhance reliability of systems. For large K-out-of-N systems, a common problem in CBM is that finding optimal maintenance solution is usually too computationally demanding due to the exponential explosion of state space and action space, when there is dependence between components in the system. In this paper, a scalable approach is proposed aiming at solving CBM problem of large-scale K-out-of-N systems with economic dependence. The approach works to find an approximate solution with an acceptable computational load. The formulation of component-wise Markov Decision Process(CW-MDP) and adjusted componentwise Markov Decision Process(ACW-MDP) are based on extended single-component actions space and distributed singlecomponent reward function so that component-level solutions are obtained and help to make system-level decision without using huge amount of computational resources. Convergence properties are analyzed in this paper to show theoretical optimality gap. Numerical study and case study are conducted to demonstrate the effectiveness and efficiency of our approaches.

Explainable Artificial Intelligence (XAI) emerged to reveal the internal mechanism of machine learning models and how the features affect the prediction outcome. Collinearity is one of the big issues that XAI methods face when identifying the most informative features in the model. Current XAI approaches assume the features in the models are independent and calculate the effect of each feature toward model prediction independently from the rest of the features. However, such assumption is not realistic in real life applications. We propose an Additive Effects of Collinearity (AEC) as a novel XAI method that aim to considers the collinearity issue when it models the effect of each feature in the model on the outcome. AEC is based on the idea of dividing multivariate models into several univariate models in order to examine their impact on each other and consequently on the outcome. The proposed method is implemented using simulated and real data to validate its efficiency comparing with the a state of arts XAI method. The results indicate that AEC is more robust and stable against the impact of collinearity when it explains AI models compared with the state of arts XAI method.

Machine learning (ML) has become extremely important in communications due to the increasing complexity of networks and due to its promises of improved performance, efficiency, security, cost reduction, and customized user experience. In wireless networks, ML is researched and used in several domains, from access networks to core networks and network services. In network security, the promises of machine learning are vast, from preventive measures to detection to response and remediation. However, machine learning requires a huge amount of resources, mainly due to the fact that ML operates on data, and data volumes are consistently rising. In this review article, we explore the aspect of resource consumption of ML techniques used for network security and provide a comprehensive review of the current state of research. Moreover, we propose a taxonomy that can be used to classify the methods through which the resource consumption can be reduced for different ML-based network security implementations. The focus of the study encompasses several key aspects pertaining to resource consumption, including energy, computing, memory, latency, bandwidth, and human resources. These resources are critical in improving the efficiency and optimizing the reliability and sustainability of network security solutions. Furthermore, based on the extensive literature review, we summarize key points regarding optimizing resource consumption in ML-based network security solutions. Finally, the challenges and future research directions for resource-efficient, ML-based network security solutions are outlined to aid in the advancement of research in this area.

Artificial intelligence techniques including machine learning models have shown success in variety of domains. This is more evident with complex models including deep learning. However, such success accompanied by vagueness around how the models work, the internal mechanism and making a decision. Explainable Artificial intelligence (XAI) emerged as a new field of research to uncover the mystery around how the complex models work. The ultimate aim of XAI is to make the complex models more transparent, trustworthy and understandable even by lay-persons with no technical background. XAI could come in different forms including heatmaps applied to images, significant concepts to the model when making a decision, informative features with tabular data, one-feature effect on the outcome, fuzzy logic rules, textual explanation through images captioning, uncertainty quantification and much more. There are several factors affect the XAI forms including the used data and the model. This paper is dedicated to review and group the current XAI methods in the literature based on the outcome form. In addition, the paper discusses the XAI groups, how they work, strengths and weaknesses. Our paper shows that although e XAI methods have been used extensively within the research context, their employability in real life problems especially in sensitive domains might be unreasonable in the current stage.

Nowadays, commercial off-the-shelf (COTS) processors are increasingly popular in embedded systems for their affordability and suitability in industrial use. However, technology downsizing has made them more susceptible to bit flips-unintended memory changes in CPU registers that can cause silent data corruption (SDC) by disrupting data and control flow. These errors can cause unpredictable behavior and reduce system reliability. In this study, we evaluate the resilience of an Autonomous Mobile Robot (AMR) by injecting bit flips into its Intel-based embedded computer. The experimental results revealed a notable vulnerability of the AMR to bit flips, reaching an SDC as high as 39.71%. While many existing methods for detecting and recovering from bit flips are effective, they often impose considerable resource demands and performance overhead, making them impractical for use in resource-constrained COTS processors. To tackle this issue, our research introduces a novel selective protection approach designed for AMR. Our methodology involves two phases. Initially, we analyzed the AMR's components based on their vulnerability to fault injection and their interaction with memory. We then applied the S-SWIFT-R protection technique to the most vulnerable components, gradually increasing the number of protected elements. The results show that selectively protecting critical components reduced SDC occurrences to below 3.1%, with a code size overhead of approximately 1.71X.

In this work, we extend the standard quasi-Helmholtz filters to high-order discretizations in the context of the boundary element method. This generalization allows preconditioning integral equations such as the electric field integral equation (EFIE) in the h-refinement regime while preserving the accuracy gain of high-order discretizations. The theoretical framework will be corroborated by numerical results that validate the effectiveness of the proposed strategy when stabilizing the refinement-dependent spectral behavior of the high-order discretized EFIE.

The first of Calderón identities indicates that the operator of the electric field integral equation (EFIE) can potentially act as a preconditioner for itself. However, the discretization of this preconditioning operator via the boundary element method (BEM) requires constructing a dual space commonly defined on a barycentric refinement of the orginal mesh. This contribution introduces a novel Calderón strategy that leverages high-order quasi-Helmholtz projectors. Unlike the standard strategies, which are based on barycentric refinements, a dual space is obtained by combinating functions of higher order. This strategy allows the development of a discretization of the Calderón identity that is suitable for arbitrary order. The well-conditioning of the proposed formulation is investigated, and numerical results are presented to corroborate the theory.

The urgent need to meet Environmental, Social, and Governance net-zero commitments and the financial risks posed by rising energy costs, are placing increasing pressure on Mobile Network Operators (MNOs) to optimise energy use. Given the significant energy consumption from Base Stations (BSs) and the fact that Fifth Generation (5G) deployments are not meeting expectations for reduced energy consumption when compared to Fourth Generation (4G), MNOs are shifting from a load-centric optimisation approach to a service-centric paradigm that balances spectral efficiency and energy efficiency. This paper introduces the Energy Sustainability Framework, which provides a standardised index to evaluate the spectral and energy efficiencies of BSs, summarised across four classes: A, B, C and D. This framework supports network auditing, benchmarking, monitoring and energy efficiency optimisation while ensuring the desired Quality of Service (QoS). By leveraging Machine Learning (ML) and Explainable Artificial Intelligence (XAI) techniques applied to live data integrated with domain expertise, the framework goes beyond deterministic computation to identify the root-cause factors influencing the BSs' energy class, thereby enabling targeted optimisation. The framework's model-agnostic nature allows it to be applied at various aggregation levels, such as location (site level), BS (cell level), or equipment type (hardware level). Notably, the implemented ML and XAI model achieved an F1-Score of 0.78 for 5G scenarios and 0.83 for 4G scenarios in predicting the Energy Sustainability Class, relying on a reduced set of indicators encompassing the most influential factors. An additional analysis, based on domain expertise, provided evidence of potential causal relationships between the class and observed variables involving channel quality, Multiple Input Multiple Output (MIMO) utilisation, network topology and energy-saving functionalities. For example, the results show that using higher order modulation coding schemes reduces the probability of a BS falling into class D by 50%. These findings provide practical insights for developing more sustainable network operations.

We present BALDUR, a novel Bayesian algorithm designed to deal with multi-modal datasets and small sample sizes in high-dimensional settings while providing explainable solutions. To do so, the proposed model combines within a common latent space the different data views to extract the relevant information to solve the classification task and prune out the irrelevant/redundant features/data views. Furthermore, to provide generalizable solutions in small sample size scenarios, BALDUR efficiently integrates dual kernels over the views with a small sample-to-feature ratio. Finally, its linear nature ensures the explainability of the model outcomes, allowing its use for biomarker identification. This model was tested over two different neurodegeneration datasets, outperforming the state-of-the-art models and detecting features aligned with markers already described in the scientific literature.

Objective: This study set out to co-design a wearable Focal Muscle Vibration (FMV) device and test its feasibility as a rehabilitation tool to reduce spasticity and improve upper limb function in stroke population. Methods: The objective was achieved through carrying out: a) a semi-structured workshop involving patients, clinicians and researchers to gather systematic user feedback, b) changes to the device design based on this feedback, c) feasibility study comparing effectiveness of 40 minutes of INTERVENTION (FMV+ conventional therapy) with CONTROL (conventional therapy alone), in reducing spasticity and improving upper limb function in 10 stroke participants, d) Quantitative (Fugl Meyer Upper Limb (FMUL), Modified Ashworth (MASh), Pittsburgh Rehabilitation Participation Scale (PRPS), Recruitment Rate and Retention Rate)) and Qualitative (verbal and written feedback), measures to assess participants perception of the wearable technology and its potential as a rehabilitation tool in reducing spasticity and improving function. Results: Participant recruitment and retention rates were high (>80%) along with high PRPS score (6/6), indicating high level of participant engagement throughout the study. FMV intervention improved both FMUE and MASh scores compared to the Control. Discussion/Conclusion: This study systematically investigated a wearable FMV device to reduce spasticity and/or improve upper limb function in stroke patients, from co-designing a device with patients and clinicians to carrying out a feasibility study. Clinical outcome measures indicate technology’s potential in reducing spasticity and improving upper limb function. Participants feedback on usability and acceptability of the technology was also highly positive, indicating its potential as a transformative rehabilitation tool.

Reconstructing the mandible after the surgical removal of diseased sections is crucial for restoring functions such as chewing, speech, and facial integrity. However, achieving consistent bone union remains a significant challenge. This study hypothesizes that optimizing donor bone cuts, specifically through improved positioning and cut plane orientation, can enhance bone union outcomes. To test this hypothesis, we present a novel automated workflow that leverages Bayesian optimization. Applied to a mandibular body defect case, this approach integrates virtual reconstruction, remeshing, and biomechanical simulation, resulting in up to a 76% increase in predicted bone union propensity. Our findings indicate that donor cut planes warrant greater emphasis in surgical planning. Furthermore, automated optimization underscores the potential of this patient-specific, data-driven method to advance bone healing in mandibular reconstruction.

Access to suitable learning platforms is crucial for motivating learners and making Science, Technology, Engineering, and Mathematics (STEM) an attractive field of study. This study analyze the potential benefit of integrating Microcontrollers into STEM learning in the schools. The responses of participants in the STEM Outreach program is analyzed to evaluate the potential impact in encouraging the rural youths in taking up the STEM fields of study in higher education. Coding on two microcontrollers, namely the runlinc E32W controller and Arduino, was introduced in ten schools, involving 447 students, of which 211 were girls and 236 were boys. The focus was on making them understand and helping them project the larger scope of STEM through developing basic applications on automation, the Internet of Things (IoT), and robotics, particularly on solving local problems through technology interventions. Pre-and post-program empirical data were collected to examine how access to and hands-on learning could influence young learners' decision to pursue STEM as their chosen area of study. To gauge the impact of the hands-on learning experience, the researchers employed a Dependent Sample T-Test, ANOVA, Pearson correlation, and linear regression analyses to predict impact. All the tests indicated a significant positive impact on the influence of appropriate learning tools in STEM interest among the students. Further, the SPSS ANN algorithm predicted the significant impact on students' perception of STEM and its future scopes.

Recursive mechanisms have become essential in bridging the gap between linear processing capabilities and the intricate demands of extended textual coherence. Yet, traditional models often struggle to retain semantic and contextual clarity across evolving dialogues and multi-layered interactions, highlighting a pressing need for adaptive context handling frameworks. Introducing Recursive Context Layering (RCL), a novel approach that recursively structures context in Large Language Models (LLMs), this study provides a transformative methodology for retaining thematic alignment and managing complex discourse across extended text. Through recursive layering, RCL establishes a cumulative context system where each layer incrementally refines semantic depth, dynamically adapting to shifts in conversational focus while mitigating the challenges of semantic drift. The RCL framework allows LLMs to seamlessly adjust to varying levels of contextual granularity, maintaining response coherence in both structured and fluid language environments. Experimental findings substantiate the efficacy of RCL, demonstrating substantial improvements in semantic accuracy, contextual consistency, and retention metrics over baseline models, marking a significant leap in LLM design for tasks requiring sustained contextual depth and responsiveness. Overall, Recursive Context Layering not only enhances contextual fidelity in language processing applications but also sets a robust foundation for recursive strategies, positioning it as a crucial advancement for models handling complex and extensive discourse.

Change detection (CD) in remote sensing (RS) imagery plays a crucial role in Earth observation tasks using satellite data. However, due to the high similarity between the foreground and background in RS images and the insufficient utilization of difference information, current deep learning (DL)based CD methods still suffer from boundary ambiguity and false changes. To address these issues, we propose an Edge-Guided Difference Enhancement Network (EGDENet). First, multi-scale feature maps are extracted from a feature extractor. To explicitly delineate the boundaries of changed objects, we introduce the Edge Enhancement Module (EEM), which captures high-frequency edge information from bi-temporal feature maps at each scale. Then, a Dual-Branch Difference Enhancement Module (DBDEM) is designed to learn the difference between the foreground and background, guided by spatial and channel information, reducing the impact of irrelevant changes on the detection results. Additionally, a Multi-Scale Difference Feature Fusion Module (MSDFFM) is employed to enhance the robustness to multi-scale features, improving the internal completeness of the detected change regions. The experimental results demonstrate that our lightweight network effectively utilizes difference information and produces more accurate change region boundaries, outperforming state-of-the-art (SOTA) methods on four public datasets.

This paper presents a new hybrid methodology for evaluating Retrieval-Augmented Generation (RAG) systems. The existing approaches provide a limited, one-dimensional assessment of RAG that lacks generalizability and is not suitable for interdisciplinary applications which cannot be extended to more than one domain. To address this, we propose a comprehensive framework that incorporates semantic metrics assessing several aspects, such as query relevance, factual accuracy, context consistency, semantic coherence, relevance, hallucination detection, and answer correctness using modern natural language processing tools. Our proposed methodology was compared to the existing state-of-the-art like llm-as-judge approach and outperformed the competitors by upto 10%. The architecture is designed for flexibility, making it applicable not only to RAG systems but also to a range of natural language generation tasks. This work extends the existing body of knowledge in both RAG systems as well as natural language generation tasks by providing a robust multidimensional evaluation approach. Robust scoring system that penalizes lower scores with Harmonic Means together with PCA, Adaptive, and Entropy weighting approaches identifies areas for improvement and provides specific retrieval and chunking method recommendations. Semantic metrics provides a low-cost alternative evaluation technique suitable for closed or offline contexts across multiple domains.