A document by Mathis Schmieder. Click on the document to view its contents.

The rise of eHealth technologies has transformed cardiac disease diagnosis, leveraging edge computing, AI, and IoT to offer critical insights into heart health. Data privacy constraints in centralized systems hinder access to large-scale ECG datasets, posing challenges for early diagnosis. While advancements in quantization and compression enable neural networks to run on edge devices, effective solutions for efficient training and inference on constrained devices are limited. To address challenges to training on the edge, we propose a mix-precision quantized DNN FPGA accelerator designed for multi-class cardiac diagnosis. Our solution achieves a top-1 test accuracy of up to 93.26% while enhancing computational efficiency, optimizing resource usage, and reducing transmission power. Our Mix-Precision Quantized FPGA Accelerator achieves up to 136x and 7.2x faster inference compared to state-of-theart Split-CNN and DCNN-Convolutional FPGA Accelerators, respectively. The accelerators offer a throughput of up to 1439.36 samples per second, latency of only 695µs, and programming logic power consumption staying below 600mW. Using hardwaresoftware co-design, our FPGA-based "Training on the Edge" approach combines software flexibility with hardware speed and improved diagnostic Top-1 test accuracy by up to 2.8% within just five training cycles, making the model more diverse to the dataset. The proposed approach also accelerates the development and reduces hardware rebuild time by a factor of (Training Cycles-1)x, ensuring efficient, sustainable ML solutions on edge devices. The source code is made available on https://github.com/shakeelakram00/Continual-Learningon-FPGAs-using-FINN

Phishing attacks are a major danger to organizations, frequently resulting in data breaches, financial setbacks, and interruptions to operations. The cultivation of phishing-resistant users is essential for both cyber resilience and the maintenance of business continuity. These attacks pose a significant threat to organizations leading to data breaches, financial losses, and operational disruptions. Employing phishing-resistant multifactor authentication (MFA), such as hardware security keys and biometric authentication, significantly strengthens security, making it harder for attackers to compromise accounts. Furthermore, having phishingresistant users gives an organization a greater capacity to quickly spot and react to phishing attempts, thereby minimizing the harm they can cause.

The unprecedented demand for seamless multimedia services necessitates novel approaches to optimize caching and computing. This paper presents an innovative framework using digital twin (DT) technology to create a dynamic, virtual representation of multimedia service infrastructure, enabling real-time optimization for user-centric applications. By leveraging DT's predictive capabilities and integrating them with advanced caching and computing strategies, the proposed system improves resource allocation, reduces latency, and enhances Quality of Experience (QoE) for users. Our simulation results demonstrate significant performance improvements over traditional approaches, establishing the potential of DT-based systems in next-generation multimedia services.

Subaerial biofilms (SABs) are microbial communities that form on surfaces exposed to both air and periodic moisture and that can adapt to harsh environmental conditions like UV radiation, and fluctuating temperatures. On the one hand they can protect built surfaces by forming a barrier against environmental stressors, on the other they can also cause deterioration through biological weathering. The balance is complex and depend on a large number of factors. Unfortunately, only a small part of the complex multiscale network of physical, chemical and biological processes is captured by existing mechanistic model; this prompts for the involvement of phenomenological models. In this work we point at the modeling advantages offered by Bayesian Networks (BNs), Causal Networks (or Structural Causal Models, SCMs) and Targeted Learning (TL). The three related frameworks can model the complex interactions between biofilms, substrates, and atmosphere by identifying cause-effect relationships. SCMs can help devise interventions able to keep the biofilm within a desirable range of conditions, furthermore they can help assess the transportability of the discovered causal models across settings; Targeted Learning helps focusing on a specific research question (such as estimating the influence of factors on a target variable; quantifying their interactions; distinguishing between direct and mediated influence) without requiring the full discovery of the joint variable distribution, and using a semiparametric approach, as free as possible from modeling biases. In this work we highlight the considerable potential of the use of those related frameworks to the study of SABs.

Eye Movement Desensitization and Reprocessing (EMDR) is a robust, eight-phase psychotherapy technique designed to help individuals reprocess distressing or traumatic memories. This method has demonstrated efficacy in reducing Post-Traumatic Stress Disorder (PTSD) symptoms, such as intrusive thoughts and avoidance behaviors, by systematically targeting maladaptive memory networks (EMDR Institute, Inc., n.d.; Linder, 2020). However, the approach still faces practical limitations, including high dropout rates, variable therapist availability, and logistical barriers—especially in underserved areas (Trauma Center NYC, n.d.; U.S. Department of Veterans Affairs, 2023). Recent advancements in Extended Reality (XR) and Artificial Intelligence (AI) offer promising solutions to these challenges. XR technologies provide immersive and customizable therapeutic environments that can replicate or simulate triggers, thereby enhancing exposure-based interventions (Inward Onward Therapy, n.d.; Vogue, n.d.). Meanwhile, AI-driven frameworks, such as the SINCLAIR Spatial Immersion Neural Computing Learning Adaptation Intelligent Reinforcement system, deploy sophisticated Machine Learning Operations (MLOps) pipelines and advanced reasoning techniques—Chain-of-Thought (CoT), Forest-of-Thought (FoT), and Trilevel Reasoning Graph (TRG)—to adaptively tailor therapy parameters in real time (Alapeti et al., 2025; Bi et al., 2024; Brown et al., 2020; Kojima et al., 2022; Zhang et al., 2024). By capturing user telemetry (e.g., eye movements, physiological cues), these AI agents propose session modifications (e.g., adjusting bilateral stimulation speed, offering grounding prompts) while preserving the authority of licensed clinicians. This paper details how XR and AI can be systematically integrated to bolster EMDR’s effectiveness in both government and commercial healthcare contexts. Compliance with U.S. Department of Veterans Affairs (VA) and Health Insurance Portability and Accountability Act (HIPAA) regulations ensures secure handling of Protected Health Information (PHI). Aligned with FDA guidance for low-risk Level I wellness products, the system remains a clinician-guided adjunct that protects patient welfare while expanding access (FedRAMP, 2023; NIST, 2002, 2020). Ultimately, the union of EMDR, XR, and AI supports more personalized treatments, reduces dropout risks, and strengthens outcomes for individuals confronting PTSD and other trauma-related disorders in both military and civilian populations.

The evaluation of large language models (LLMs) has traditionally relied on static benchmarks that prioritize performance metrics such as accuracy, precision, BLEU and standardized tests. However, these metrics frequently fail to capture the deeper reasoning, consistency, and real-world applicability of these models. This paper critically examines the limitations of the current benchmarks and preliminarily introduces a practical, context-sensitive framework to qualitatively evaluate LLMs. Through two case studies, which include one for medical diagnosis and the other for traffic engineering, we highlight key shortcomings in existing evaluation approaches. Despite achieving accuracy of 92% in diagnosing diseases and solving multi-step reasoning tasks in the traffic engineering case with 62% accuracy, the evaluated LLM displayed critical gaps, such as overlooking ethical considerations, practical constraints, and failing to maintain consistency across varied prompts. Expert evaluations further revealed issues in reasoning transparency and technical feasibility. This might raise questions about trust in the applicability of LLM systems. By emphasizing domain-specific tasks, expert involvement, and consistency analysis, this study advocates for a paradigm shift in how we benchmark LLMs. The findings underline the urgent need for more nuanced evaluation frameworks to ensure LLMs can reliably support complex, real-world decision-making.

This study explores the representation of R1 and R2 universities on social media platforms through topic modeling and sentiment analysis. Using Wikipedia content as a dataset, the research applies Latent Dirichlet Allocation (LDA) to uncover key themes and employs computational tools, such as VADER and TextBlob, to analyze sentiment trends. The results reveal three primary topics: educational infrastructure, academic and student development, and campus life. Sentiment analysis indicates that the majority of content reflects neutral or positive tones, emphasizing the professional and constructive nature of the discourse. The findings offer insights into how universities project their image online and demonstrate the importance of data preprocessing in shaping analytical outcomes. This study provides a foundation for leveraging social media analysis in university branding and public engagement strategies.

Efficient processing of large data volumes at low power is a critical challenge in modern computing systems. Spiking Neural Networks (SNNs), inspired by biological spikebased communication, offer a promising solution through energyefficient neuromorphic architectures. Unlike traditional neural networks, SNNs utilize spike-timing-dependent plasticity (STDP) learning, reducing computational complexity and power consumption. Prior implementations often rely on external computational platforms or memristive devices, which introduce fabrication and integration challenges. This work presents the design and implementation of an on-chip SNN architecture using SCL 180nm CMOS technology. The architecture features Leaky Integrate-and-Fire (LIF) neurons and a memristive synapse circuit that incorporates STDP learning for efficient pattern recognition. Optimizations, such as removing current mirrors and reset logic in the neuron circuit, minimize area and power requirements. A Verilog-A designed summing amplifier processes synaptic outputs, while a Winner-Takes-All mechanism ensures accurate classification of input patterns, including characters and directional signs. The single-layer network demonstrates lowlatency performance and high efficiency in recognizing simple patterns. Simulation results validate the robustness of the architecture, paving the way for future enhancements to include hidden layers for handling complex patterns and integrating advanced technologies to expand the scope of neuromorphic computing applications.

Magnetic flux tubes and magnetic monopoles are interconnected concepts in theoretical physics, particularly in the concept of gauge theories and Superconductivity.This paper discusses this concept from the view of superconductivity.

The rapid evolution of ransomware poses an increasing threat to information security, with malicious actors deploying advanced techniques to evade traditional detection systems. Introducing a method for ransomware detection that leverages Gaussian algorithms for classifying low-level file system I/O Request Packet (IRP) patterns offers a novel and statistically grounded approach to identify deviations in system behavior. The methodology captures abnormal IRP sequences that signify ransomware activity, outperforming conventional signaturebased and heuristic detection techniques, especially when dealing with zero-day ransomware variants. A controlled experimental environment demonstrated the effectiveness of this approach, achieving high detection rates and minimizing false positives. Additionally, the model's adaptability to dynamic ransomware behavior positions it as a robust solution for environments where early intervention is critical. Results indicate that the Gaussian algorithm-based model provides a scalable and efficient method for real-time detection of ransomware, significantly enhancing the capabilities of modern cybersecurity frameworks.

Single sign-on authentications ease the challenges of creating new pathways into different software and databases for users. However, with the growing use of web bots and malicious actors looking to take advantage of sensitive data, it becomes imperative that a well-meaning authentication method of filtering thoughtful users be utilized. This article presents a large-language model powered mechanism and structure to filter users to be human, non-malicious, and conscious of their membership of accounts, databases, and other software.

The integration of terrestrial Internet of Things (IoT) and Non-Terrestrial Networks (NTNs) is a fundamental enabler of ubiquitous connectivity in future sixth-generation (6G) wireless networks. This survey provides a comprehensive assessment of state-of-the-art advancements, key challenges, and emerging industrial applications in IoT-NTN convergence, with a focus on edge intelligence, energy-efficient communication protocols, and direct-to-device (D2D) NTN access. Through a systematic analysis of over 30 seminal studies, we identify critical research gaps in scalability, latency minimization, and interoperability within standardized frameworks. Furthermore, we outline future research trajectories, emphasizing AI-driven network orchestration, quantum-secured NTN links, and dynamic spectrum management to enhance network resilience, security, and spectral efficiency in next-generation NTN-integrated IoT architectures.

Cellular heterogeneity is one of the main hallmarks of cancer, referring to the coexistence of different phenotypes with very distinct biological behaviors in single isolates. Automatically detecting single-cell heterogeneity is therefore critical, and can provide important information on cancer initiation, tumour aggressiveness or drug-related resistance. Here we introduce He2Cl, a novel machinelearning algorithm able to detect single cell heterogeneity from the morpho-dynamic analysis of 2D cell cultures. Built on label-free quantitative time-lapse imaging, He2Cl performs a 2-step clustering algorithm for unsupervised classification of sub-phenotypes within a cell population. He2Cl outperforms state-of-the-art clustering models and allows the simultaneous classification of thousands of cell trajectories from their birth to their division, paving the way towards AI-enhanced microscopy for live-cell analysis.

Autonomous Vehicle decisions rely on multimodal prediction models that account for multiple route options and the inherent uncertainty in human behavior. However, models can suffer from mode collapse, where only the most likely mode is predicted, posing significant safety risks. While existing methods employ various strategies to generate diverse predictions, they often overlook the diversity in interaction modes among agents. Additionally, traditional metrics for evaluating prediction models are dataset-dependent and do not evaluate inter-agent interactions quantitatively. To our knowledge, none of the existing metrics explicitly evaluates mode collapse. In this paper, we propose a novel evaluation framework that assesses mode collapse in joint trajectory predictions, focusing on safetycritical interactions. We introduce metrics for mode collapse, mode correctness, and coverage, emphasizing the sequential dimension of predictions. By testing four multi-agent trajectory prediction models, we demonstrate that mode collapse indeed happens. When looking at the sequential dimension, although prediction accuracy improves closer to interaction events, there are still cases where the models are unable to predict the correct interaction mode even just before the interaction mode becomes inevitable. We hope that our framework can help researchers gain new insights and advance the development of more consistent and accurate prediction models, thus enhancing the safety of autonomous driving systems.

The evolution of cyber threats targeting sensitive data and systems has necessitated the development of increasingly advanced detection techniques to counteract their sophistication and adaptability. Dynamic Encrypted Payload Analysis introduces a groundbreaking methodology, addressing limitations in existing approaches through real-time monitoring and pattern recognition of malicious encryption behaviors. Through leveraging entropy analysis and machine learning models, the framework achieves superior accuracy and scalability while reducing false positives, making it highly effective against adaptive ransomware strategies. Comprehensive experimental evaluations have validated its performance across diverse environments, showcasing significant improvements in detection speed, resource efficiency, and adaptability to emerging threats. The findings demonstrate the transformative potential of the proposed framework, providing a resilient, scalable, and efficient solution for securing digital infrastructures against evolving ransomware attacks.

In the age of digital transformation, the rise of Big Data has fundamentally altered how organizations store, process, and utilize information. This whitepaper provides a comprehensive analysis comparing Big Data with traditional data systems, data warehousing, business intelligence (BI), artificial intelligence (AI), data science, and NoSQL databases. By exploring key differentiators such as volume, variety, velocity, and processing capabilities, this paper aims to shed light on how Big Data has reshaped modern technology infrastructures and its role in advancing analytics, decision-making, and operational efficiency.

Harnessing the potential of smart grid networks relies on the efficient and secure communication between distributed energy resources (DERs) and the energy management system (EMS), particularly under variable channel conditions and adversarial interference. This interference is further exacerbated with the advent of artificial intelligence (AI)-driven adversarial devices capable of adaptively disrupting communication in real time. To address these challenges, in this study, we formulate the distributed channel access problem, incorporating dynamic channels and intelligent adversarial interference as a partially observable Markov game (POMG). Specifically, we propose a distributed framework based on a centralized training and distributed execution (CTDE) multi-agent reinforcement learning (MARL), enabling DERs to autonomously adapt to dynamic channels and mitigate intelligent interference using only local observations, i.e., without direct information sharing among DERs. Our simulation results indicate that, by integrating an advanced policy evaluation technique and a tailored utility maximization strategy, DERs can collaboratively optimize their transmission decisions, improving the network's aggregate packet success rate (APSR) and resilience. Additionally, the proposed framework outperforms existing methods, ensuring robust and scalable communication in smart grids under diverse conditions.