Siste publikasjoner ved Institutt for informasjons- og kommunikasjonsteknologi

Machine learning (ML) offers a scalable approach for dynamically predicting learning styles (LS) by analyzing behavioral data from Moodle interaction logs, reducing reliance on traditional questionnaires. However, most studies rely on a single LS framework. This study proposes a hybrid cluster–classifier architecture for LS prediction and a systematic cross-framework comparison within a Moodle environment. The hybrid ML approach combines K-means clustering to derive pseudo-labels with supervised ensemble models to predict LSs based on behavioral interaction data. Decision Trees, Random Forests, Gradient Boosting, XGBoost, and Neural Networks were tuned and evaluated using stratified five-fold cross-validation. The approach evaluates prediction performance and examines classification correspondence between two complementary models: the Felder–Silverman learning styles model (FSLSM) and VRK (visual–read/write–kinesthetic, a subset of VARK). High reproducibility of behaviour-derived clusters was achieved for three of the four FSLSM dimensions (perception, input, and understanding), as well as for VRK styles; however, the FSLSM processing dimension showed weaker cluster separation. Cross-framework feature-importance analysis revealed both overlapping and framework-specific behavioral indicators, reflecting convergence and divergence in how the frameworks captured behavior. This triangulation provides a nuanced and complementary understanding of learning preferences compared with any single-model approach. The findings demonstrate the feasibility of hybrid LS prediction and cross-framework comparison using Moodle behavioral data and provide empirical evidence supporting behaviour-based learning preference modelling. The proposed approach provides a foundation for behavior-aware, non-prescriptive personalization and informs the design of more inclusive e-learning environments.

In a multi-agent system, platoon vehicles receive a huge collection regarding autonomous models coordinating and their actions to improve traffic flow, lower fuel consumption, and boost safety. This paper examines the distributed consensus control problem for heterogeneous multi–agent systems (MASs) containing both first-order and second-order agents, under constrained network communication resources. Secondly, a novel event–triggered approach is proposed to tackle the problems of information transmission restrictions and bandwidth contention. Unlike conventional state-independent triggering methods, the proposed trigger condition depends on both the agent’s own state update error and the information mismatches between neighboring agents, enabling a balanced trade–off between control performance and communication reduction. By ensuring that the transmission interval always exceeds one sampling period, the event-triggered technique significantly reduces network bandwidth usage. It determines the next transmission time for both position and velocity information. Thirdly, sufficient criteria for reaching asymptotic consensus are determined using Lyapunov stability theory and Kronecker product features. The trigger parameters and controller gains are obtained by solving linear matrix inequalities (LMIs). The distributed event–triggering mechanism and the consensus control protocol are integrated into a co-design framework. The effectiveness of the suggested strategy in conserving network resources without sacrificing system stability is validated by simulation results for a heterogeneous MAS with two second–order and two first–order agents, which show that all agent states converge to common values while the number of information transmissions is significantly reduced compared to periodic sampling.

This article proposes a progressive feature selection framework for millimeter-wave (mmWave) radar-based wall material classification. Raw complex baseband I–Q data are collected using a TI AWR1843 radar over multiple days from eight indoor surface types: transparent glass wall, smooth glass board, cement wall, metal door, vinyl floor, olefin carpet floor, ceramic tile, and concrete wall. We introduce an incremental-data decremental-feature learning (IDDFL) strategy that incrementally aggregates daywise training data while progressively pruning low-contribution features using a posthoc explainer. A preprocessing module extracts 100 handcrafted descriptors covering time-domain, frequency-domain, time–frequency, and radar-specific characteristics from the acquired I–Q measurements. Feature relevance is quantified via SHapley additive exPlanations (SHAP) analysis on a LightGBM (LGBM) classifier, and the least discriminative features are removed across days. The proposed IDDFL process reduces the feature set from 100 to 30 while maintaining high classification performance. The proposed model achieves a final test accuracy of 79.80% with an F1 -score of 0.79 , demonstrating improved computational efficiency with minimal performance degradation. The final 30-feature LGBM model outperforms nine state-of-the-art benchmark methods reported in the literature and is deployed on edge platforms (Raspberry Pi 4 and Rock Pi 4 C+) to demonstrate real-time applicability. The proposed feature selection and edge-ready pipeline is well-suited for building inspection, autonomous navigation, and security applications.

The evolution of the service-based architecture (SBA) requires cloud-native core networks, and its open and modular design nature makes Kubernetes the most eminent platform for SBA implementation. One critical-yet-underexplored issue when deploying the SBA based on Kubernetes is the accumulation of residual artifacts that persist and degrade observability and stability inside Kubernetes clusters. In this paper, we propose residual artifact governance (RAG), a non-intrusive add-on module to the SBA that enforces bounded garbage collection (GC) policies without modifying the underlying architecture. The module follows a master–agent structure: a GC-master derives data-driven machine learning-enabled cleanup policies using the GC data provided by GC-agents, while GC-agents co-located at each network function execute deterministic, auditable cleanup actions. As a proof-of-concept demonstration, we implement a Kubernetes-based prototype to validate the operability and feasibility of RAG. While a machine learning predictor located at the GC-master provides early warning signals on time-to-collapse caused by the accumulation of residual artifacts, periodic fractional cleanup actions at GC-agents keep residual peaks bounded under both nominal and stress leak regimes. Results reveal that the accumulation of residual artifacts can be measured, predicted, and controlled using Kubernetes-native mechanisms through our RAG enhancement while remaining compatible with 3GPP-aligned observability and stability assurance principles.

The Tsetlin machine (TM) is an interpretable machine learning (ML) algorithm that requires relatively modest computational resources. While most ML models use continuous data representations, TM operates on binary inputs. Binarization of continuous data via brute-force thresholding has shown promising accuracy but can be computationally expensive and hinder generalization. Moreover, traditional binarization approaches, such as mean, median, and fixed thresholds, yield single-bit representations. Multi-bit representations, such as quantile, histogram, and standard binarization, are more expressive, but typically require manual definition of the number of threshold levels or the number of bits per feature. However, not all features require the same number of bits, nor do they contribute equally to prediction. To address this, we propose a simple and adaptive binarization framework that automatically determines how each feature should be represented based on its data characteristics. Instead of relying on fixed thresholds. This method identifies meaningful groups within each feature and encodes them accordingly, resulting in a more informative and balanced representation. In this paper, we propose a significance-aware adaptive binarization framework for TM learning based on k-means clustering. Each feature is individually analyzed using the Silhouette score to determine its optimal clustering structure. Features are then one-hot encoded according to their cluster partitions, resulting in a significance-aware binarized representation. All the binarized features are concatenated to form the final input for the TM. To ensure scalability, the framework includes an efficient mechanism to handle large datasets while maintaining consistency in the encoding process. To reduce the complexity of clustering, we propose a chunk-based centroid-to-centroid association mechanism, efficiently solved using the Hungarian algorithm. The proposed method is evaluated on multiple datasets and benchmarked against traditional single-bit and multi-bit binarization techniques. Moreover, we extend this analogy to continual feature binarization.

This study presents a Systematic Literature Review (SLR) of methods and challenges related to distinguishing cybersecurity attacks and safety faults in Distributed Energy Resources (DER)-rich smart grids. The increasing integration of DERs has improved grid flexibility and sustainability, but has also introduced greater operational complexity and expanded the attack surface of smart grid infrastructures. In such environments, cyber-attacks and safety faults may produce similar anomalies, making accurate distinction a critical requirement for resilient and secure grid operation. This review examines the literature on anomaly detection and event classification in DER-rich smart grids towards distinguishing these anomalies and events, with a particular focus on approaches that could support the differentiation of malicious and non-malicious events. It analyzes the main categories of data used in the literature, including sensor data, smart meter data, network traffic, phasor measurements, and weather-related information. It also reviews machine learning and artificial intelligence techniques, including supervised, unsupervised, and deep learning approaches, and discusses their applicability, strengths, and limitations within different contexts. Furthermore, the review synthesizes evaluation practices, operational application domains, and key open challenges, including limited dataset realism, data quality issues, explainability, scalability, and model generalization. Existing testbeds are highlighted as essential for reproducing realistic grid conditions and validating approaches; however, in their current state, no testbed framework is directly applicable to distinguishing cybersecurity attacks from operational safety faults in an informed manner. Importantly, only 6 of the 25 included primary studies explicitly attempt to distinguish cyber-attacks from physical safety faults within a single evaluation setting. The review is therefore explicitly positioned as a combined systematic mapping and gap analysis: rather than demonstrating a mature solution space, it documents how sparsely the central distinction problem is addressed, characterizes the few existing attempts in depth, and derives a research agenda from this gap. Overall, the review identifies major research gaps and outlines directions for developing more safe and secure solutions for DER-rich smart grids.

Advances in Quantitative Structure Activity Relationship (QSAR) are led by two core paradigms, (1) descriptor engineering, where complex fixed-length vectors of compounds are generated and conventional ML methods are applied to those representations and (2) graphical chemical inputs (e.g., Simplified Molecular Input Line Entry System (SMILES), 2D-graph) being provided to deep learning neural network (NN) models, which construct their own internal representations of molecules and learn iteratively over them. Here we present the Tsetlin Machine (TM)─which combines the accuracy and easy-use of existing rule-based QSAR ML methods (e.g., RF and XGBoost), the iterative learning aspect of NN algorithms and its intrinsic interpretability. The TM uses teams of finite-state automata which capture frequent patterns as propositional logic (clauses) via reinforcement learning. The benchmarking pipeline presented here demonstrates that TM-QSAR coupled with ECFP4 descriptors frequently performs better than existing rule-based QSAR methods for ROC-AUC, PRC-AUC and PPV, with a high capacity for interscaffold generalization. However, due to the binary nature of TM-QSAR, performance it is currently limited when descretised continuous descriptors are used. TM-QSAR demonstrated particularly impressive classification scores for MOR (ROC-AUC = 0.87, PRC-AUC = 0.77) and CYPA4 (ROC-AUC = 0.92, PRC-AUC = 0.63), when compared to RF and XGBoost. Using TM in combination with substructural fingerprinting descriptors allows for an interpretability suite which can be extracted directly from clauses. Here we detail molecule property maps (TM-MPM) to view atom-wise TM-QSAR bioactivity contributions for single molecules and closed-form WAC scores (Weights × Activations × Clauses) for descriptor-wise contributions to regions of predicted chemical space. These methods show strong alignment of TM-QSAR interpretations to known ligand–protein interactions of the MOR target and gives nonlinear, conditional interpretations for greater predicted bioactivity. Given this combination of accuracy, computational efficiency and interpretability, we provide a basis for TM-QSAR to be explored as a standard methodology in virtual screening toolkits.

The rapid advancements in autonomous vehicles (AVs) highlight the need for robust perception and planning systems to navigate complex traffic scenarios. AVs rely on congestion state estimation to assess real-time traffic, optimize routes, and navigate safely in dynamic environments. Accurate congestion estimation faces challenges like dynamic conditions and scaling to process large urban traffic data. Existing methods use spatio-temporal graph models with LiDAR or cameras. While LiDAR offers high accuracy, its cost makes camera-based systems a viable alternative. Single-camera setups are limited by narrow fields of view and incomplete coverage, highlighting the need for multi-camera approaches that enhance coverage and scalability. In this work, an ego-centric perspective on traffic scene understanding is adopted, where congestion is defined relative to the autonomous vehicle’s immediate surroundings rather than global traffic flow. Grounded in this perspective, an end-to-end Autonomous Vehicle Congestion State Classification Network (AVCSCNet) is proposed using the past-and-future reasoning for tracking (PF-Track) backbone to generate bird’s-eye view (BEV) visualizations of multi-camera inputs. BEV data is transformed into 2D spatio-temporal graphs, with the ego vehicle and surrounding agents as nodes and edges representing relationships. Further, GNNs are employed for spatial feature extraction, followed by LSTMs for traffic state classification into: Congested, Neutral, and Free-flow. In addition, we introduce the Autonomous Vehicles Congestion 3 (AVCong3) dataset, featuring manually annotated spatio-temporal graphs generated using multi-camera data and three traffic classes. Extensive experiments on AVCong3 and PREVENTION datasets demonstrate the effectiveness and scalability of the AVCSCNet framework. In particular, AVCSCNet achieves 86.6% congestion classification accuracy using six cameras, significantly improving from 68.3% with a single camera setup, while also demonstrating promising generalization on the PREVENTION dataset with 78.05% accuracy.

Cyberbullying detection in Chinese social media presents unique linguistic and computational challenges due to the widespread use of homophones, tonal variations, and adversarial text perturbations that obscure abusive intent. Conventional character- or semantic-based models often struggle to identify such phonetic disguises, limiting robustness in real-world applications. To address this issue, this paper proposes a Pinyin-based adversarial learning framework that encodes Chinese characters into structured phonetic vector representations, explicitly preserving pronunciation features. A phonetic dictionary optimization strategy consolidates homophonous entries using a Minkowski-distance threshold, reducing lexical redundancy by 33% while maintaining discriminative capacity. A similarity-driven perturbation correction mechanism further mitigates homophone-based adversarial manipulation through threshold-controlled normalization. The phonetic representations are integrated with semantic embeddings and evaluated across RNN, CNN, and BERT architectures to ensure architecture-independent validation. Experimental results on the Weibo_Senti_100K dataset demonstrate that the BERT + Pinyin model achieves strong classification performance (Accuracy 0.98, F1 0.98, AUC 0.98) and maintains high robustness under adversarial perturbations, outperforming semantic-only transformer baselines. These findings indicate that incorporating structured phonetic modeling enhances resilience against pronunciation-preserving obfuscation and provides a practical solution for robust Chinese cyberbullying detection.

This study presents an agent-based simulation framework for analysing social response during crisis situations, using a fire evacuation scenario as a controlled experimental setting in which the hazard is represented through the spread of smoke. Individuals are modelled as agents with different behavioural strategies, including altruistic, selfish, herding, and rational, interacting in a confined environment affected by smoke, crowding, and exit-capacity constraints. Agents update their behaviour based on local conditions, internal states such as panic and health, and interactions with nearby agents. Three controlled experiments explores how social behaviour, evacuation infrastructure, and hazard severity influences outcomes. The results reveal a non-linear relationship between altruism and survival, identify exit capacity as a critical structural constraint, and indicate that faster smoke spread increases panic, injury rates, and evacuation time. Overall, the findings demonstrate how simple individual decision rules generate complex collective evacuation dynamics under emergency evacuation.

The integration of artificial intelligence (AI) in education has accelerated globally, yet limited attention has been given to teachers’ acceptance of AI-powered applications, particularly in developing-country secondary school contexts. This study examines Mathematics and Science secondary school teachers in Rwanda, focusing on their acceptance of AI-powered applications and the contextual factors shaping adoption. A convergent mixed-methods design was employed, involving a survey of 120 teachers and qualitative interviews with 18 participants. Quantitative results showed that teachers reported high perceived usefulness (M=4.69) and moderately high ease of use (M=4.12), with perceived usefulness emerging as the strongest predictor of acceptance (β=0.609, p<0.001). Technology self-efficacy, social influence, and administrative support further facilitated adoption, while workload concerns and infrastructural limitations negatively affected use. Qualitative findings reinforced these trends, revealing that AI applications enhanced instructional planning, supported inquiry-based learning, and increased participation among shy or low-performing students. However, teachers expressed concerns about internet instability, insufficient training, and the potential erosion of professional judgment in assessment. Therefore, the study demonstrates that while teachers view AI as a valuable tool for improving instructional efficiency and student learning, meaningful adoption requires reliable infrastructure, structured professional development, and clear policy guidance. The findings contribute to extending UTAUT in low-resource settings and offer practical insights for policymakers, school leaders, and developers seeking to promote context-appropriate AI integration in education.

Monitoring water quality in water utilities via multiple distributed sensors is critical for detecting diverse pollution events. Multivariate water-quality time series, generated by such water quality sensors, often exhibit overlapping seasonal patterns, long-term trends, and irregular disturbances, making anomaly detection challenging without extensive labeled data. In this work, we present an anomaly detection approach based on unsupervised machine learning (ML) using residual feature embeddings and multivariate clustering. To extract anomalies, for each sensor, we compute candidate seasonal periods and fit a polynomial trend, yielding residual vectors that serve as feature embeddings for Isolation Forest (IF), an unsupervised ML approach. IF was used to identify anomalous observations through residual embeddings, and K-Means clustering was used to group residual vectors into similar water-quality events. To aid interpretability, the clusters were further analyzed statistically to identify the potential signatures of anomalous events. Validation is performed on data annotated by domain experts. This approach enables robust unsupervised anomaly detection by effectively handling multiple overlapping seasonality and trends that might otherwise obscure anomalies. An experimental evaluation of multivariate water-quality datasets demonstrates that the proposed method successfully identifies unusual events and provides valuable insights for water utilities.

Abstract Reliable time-series forecasting under rapidly changing conditions remains a critical challenge across many domains, particularly in healthcare, where physiological signals are inherently dynamic and uncertain. Blood glucose level prediction exemplifies this challenge, as accurate and timely forecasts are essential for effective diabetes management, yet traditional approaches rely on fixed prediction horizons and single-point estimates, which may yield unreliable decisions under rapidly changing physiological conditions. In this work, we propose a novel approach for adaptive horizon selection, applied to BGL prediction, employing a deep learning model. It employs evidential learning-based uncertainty quantification that decompose uncertainty into epistemic and aleatoric. Each set of models is trained to predict blood glucose levels at different future time steps, each providing both a point prediction and an associated uncertainty measure. At inference time, it dynamically balances predictive accuracy and reliability by selecting the longest horizon whose predicted uncertainty remains below a predefined threshold. This enables confidence-based horizon selection, using longer prediction horizons during stable periods and switching to shorter horizons when uncertainty signals critical glucose events requiring immediate intervention. This uncertainty-aware prediction approach promotes transparency by exposing confidence levels alongside predictions. Applicable to time-series forecasting tasks broadly, the proposed framework demonstrates encouraging potential, and when applied to BGL prediction as a representative clinical case, shows particular promise for supporting glycemic management through calibrated uncertainty estimation, offering a more transparent and interpretable alternative to fixed-horizon models toward trustworthy decision support in diabetes care.

Interpretable machine learning is increasingly sought in next-generation cellular networks, where trust, accountability, and robustness are essential and jamming remains a persistent threat. Federated learning enables collaborative jamming detection without sharing raw in-phase and quadrature data, while Tsetlin Machines provide a logic-based option for transparent, resource-constrained deployments. However, the security implications of interpretability in federated wireless environments remain poorly understood. We introduce a backdoor threat model for interpretable federated jamming detectors, showing how Tsetlin Machine clause structure can be exploited by a compromised client to induce targeted detection failures. In a federated framework with weighted Coalesced Tsetlin Machines and Krum aggregation, we analyze clause-based vulnerability to spectrogram-domain trigger perturbations and formulate a multi-objective problem balancing trigger effectiveness, model deviation, and aggregation selection. We further show that clause-informed trigger initialization lowers adversarial optimization effort, revealing a tension between transparency and robustness. Evaluation on 60,000 spectrograms from a 5G NR synchronization-signal testbed shows that optimized triggers can exceed 70% trial-level triggered false-negative rates in optimizer benchmarks, reaching approximately 90% in best-case configurations; supplemental multi-round federated evaluation shows that trial-level and global triggered false-negative rate (TFR) can diverge substantially (gaps from -31.4 to +3.0 percentage points across evaluated partitions), with no sustained global TFR≥70% after the last malicious selection round. These results highlight the need for defense-aware design of explainable models in next-generation networks.

Stator inter-turn short circuit faults in shipboard synchronous generators (SGs) can rapidly escalate from minor insulation degradation into catastrophic failures, threatening the operational continuity and safety of electrified vessels. Therefore, early fault severity estimation of SGs is essential to prevent extensive downtime and ensure the resilience of shipboard power systems. However, the conventional data-driven approaches typically require labeled faulty data, which are rarely obtainable in practical maritime operations. To address this limitation, this article proposes a machine learning framework that enables early fault severity estimation without the need for experimental fault samples for training. The proposed method couples two neural networks through a physics-guided cycle-consistency (PGCC) training strategy. This approach facilitates both high-fidelity synthetic data generation and domain adaptation from analytical simulations to experimental measurements. The framework is validated on an in-house 7.5 kW test bench, demonstrating that the proposed method achieves high-accuracy prediction of unseen experimental fault severities. Ablation studies confirm the contribution of each network component and the PGCC training strategy. Furthermore, a comparative study demonstrates that the proposed method outperforms the tested domain adaptation and generative adversarial networks-based methods, particularly under limited training data.

In a Software-Defined Vehicular Network (SDVN), the operation of vehicular networks can be improved due to optimizations at the controller driven by a global network view. However, such performance can be drastically affected if the network topology is poisoned by Link Discovery Attacks (LDAs), causing packet loss, exhausting the network, and enabling eavesdropping. Recent solutions for LDA mitigation are centralized with no fault tolerance and assume a high trust in the controller in the LD process. Moreover, they struggle to mitigate the attack for dynamic SDVN LD owing to frequent topology changes with non-adaptive link rediscovery. Other key shortcomings include (i) lack of distributed node behavioral monitoring and blacklisting, (ii) high communication interception, (iii) lack of authenticity and integrity, and (iv) relying on single observation without distributed consensus. In order to respond to this research gap, we first demonstrate 5 LDAs that can exist in a Zero-Trust (ZT) standard SDVN architecture by exploiting the vulnerabilities in location tracking and the LD service. Our main contribution is to propose a novel framework to detect LDA by using Q-learning to derive link states by considering the link’s suspicious behavior, packet delivery, and latency and with a location-driven adaptive LD technique. Attack mitigation is achieved through (i) cycle- and non-cycle-based digital signatures with mixed post-quantum and classical cryptography, (ii) Zero-Knowledge Proofs (ZKPs) of Hash Message Authentication Codes (HMACs) stored in the blockchain, (iii) time-limited session-based HMACs, (iv) packet limiting with signature matching, (v) location verification and behavioral monitoring, and (vi) symmetric key deriving using Diffie Hellman. Moreover, a permissioned blockchain is utilized by using 3 Smart Contracts (SCs). The first SC handles node registration using a challenge-response mechanism and subsequent node authentication using signatures, ZKPs, attributes, and encrypted HMACs. The second SC is used to maintain non-tamperable node reputation and location scores by monitoring node behavior and maintaining endorser trust scores for weighted reputation updates for consensus-based node blacklisting. Furthermore, a third SC is used to verify the discovered link state and penalize the controller’s malicious behavior using the trust-score-weighted blockchain endorsement policy-driven consensus of the broadcasted HELLO packet and packet-out received nodes to further evaluate the controller’s trust and switch it under low trust. Two operational modes: BQLLDA-H (High-performing), which offers maximum security, and BQLLDA-L (Lightweight), optimized for resource-constrained deployment by selectively disabling ZKP authentication, reverse LLDP verification, and post-quantum primitives, while enabling unicasting key pair, rule-based adaptation, and single-stage authentication. The proposed BQLLDA was simulated using NS-3 and Hyperledger Fabric and is effective in attack mitigation, as the combined attack evaluations for the proposed BQLLDA-H resulted in (i) the best Packet Delivery Ratio (PDR) (97%–98% under 80% attacks), (ii) zero Packet Interception Ratio (PIR), (iii) the best Matthews Correlation Coefficient (MCC) (0.95-0.66 under 80% attacks), (iv) competitive channel utilization (14%–37%), and (v) latency (24-170 ms), albeit with higher computational complexity (46-323 ms) compared to existing approaches. BQLLDA-L achieves a complexity of 18-95 ms (60%–70% reduction) with graceful degradation to 88.7% PDR, 8.4% PIR, and 0.52 MCC at 80 percent attacks.

Status update presentation of the SURE-AI Centre to the RCN Portfolio Board for Enabling Technologies

Short presentation of SURE-AI Centre for Sustainable, Risk-averse and Ethical AI

Presentation of the SURE-AI Centre for Sustainable, Risk-averse and Ethical AI and possibilities for collaboration

Bruken av KI vekkjer sterke kjensler og store overskrifter både i form av glede og forrakt, og det verkar som det er veldig ulike oppfatningar om når det er greitt og ikkje å bruke det, så vi vil med spørjetimen få KI-ekspertane sjølv til å hjelpe oss med å vurdere dette.

Spør Ekko: Når er det greit å bruke KI?

Paneldebatt om hvordan kunstig intelligens påvirker utdanning, yrkesvalg og arbeidsliv.

Thermo-mechanical simulations are central to predicting solidification behaviour, hot‑spot formation, defect evolution and distortions in shape‑casting processes. However, finite‑element–based modelling remains computationally demanding due to mesh‑generation effort, stiff non-linear heat‑transfer equations, and the need to solve PDE residuals at a very high spatial resolution. These challenges limit the use of FEM solvers in rapid design iteration, optimization loops, and real‑time process control. Physics‑informed machine learning (PIML) provides a promising route to alleviate these constraints by embedding governing thermal physics directly into neural‑network training [1]. In this work, we develop a hybrid physics‑informed neural network (PINN) framework tailored to simulating low‑pressure die casting (LPDC) processes used in the manufacture of structural automotive components at Aludyne in Norway. Traditionally run with Finite Element Methods (FEM) the main objective of the simulations is to accurately predict the final shape of the automotive component after casting and compare it to the intended CAD geometry. During each LPDC production cycle, the mould is repeatedly heated, causing it to deform under thermal loading. Consequently, the liquid aluminium solidifies inside a mould that is already distorted, producing a part in an unintended shape. Predicting and correcting these thermally driven shape changes requires a detailed numerical model, and the complexity involved makes vanilla PINNs inadequate [2]. The hybrid PINN formulation incorporates the heat‑transfer Partial Differential Equations, relevant boundary conditions, and temperature‑dependent material behavior along with data from a simplified FEM model into a unified loss function. This enables the model to approximate transient temperature fields and solidification‑front motion while predicting distortions and stresses in the mould and the cast geometry. Predictions from the hybrid PINN framework are validated against FEM results to assess the performance of the proposed approach, and future work will extend this validation to simulations of an actual production part.

«Barnehagens digitale praksis skal bidra til barns lek, kreativitet og læring», heter det i rammeplanen. Hvordan møter man det det på en sikker, lovlig og etisk forsvarlig måte?

Warped Gaussian processes (GPs) handle non-Gaussian observations by mapping them into a latent standard GP via a parametric transformation called warping. Existing online variants either optimize the warping parameters periodically rather than sequentially, or sacrifice analytical tractability and computational efficiency for higher model expressivity. In this paper, we avoid these compromises. To achieve this, we first show that the gradient of the instantaneous negative log-likelihood of a warped GP admits an exact recursive computation. Based on this result, we propose a novel online method for warped GPs that jointly updates the latent GP moments and optimizes the warping parameters. Finally, we validate our method on an existing benchmark adapted to an online formulation.

Smarte byer er i økende grad avhengige av heterogene sensormodaliteter, som trafikkvideostrømmer og tekstlige beskrivelser, for ˚a støtte intelligente transportsystemer (ITS) og urban beslutningsstøtte. Gjenfinning av trafikkscener basert på naturlig språk gjør det mulig for operatører ˚a identifisere kjøretøy eller hendelser ved hjelp av menneskelig forst˚aelige forespørsler, men forblir utfordrende på grunn av modalitetenes heterogenitet, domene-spesifikk tvetydighet og begrenset robusthet i både unimodale og svakt justerte multimodale tilnærminger. Denne avhandlingen adresserer utfordringen ved å konstruere et enhetlig multimodalt gjenfinningsrammeverk for urbane trafikkscener som systematisk integrerer representasjonslæring, innebyggingsjustering og strukturert fusjon på beslutningsnivå. Hovedbidraget er å vise at robust gjenfinningsytelse krever koordinert systemdesign på tvers av komponenter, snarere enn isolerte modellforbedringer. Forskningen utvikles gjennom stadig mer raffinerte gjenfinningsarkitekturer, med et modulært baseline-system som kombinerer konvolusjonsbaserte visuelle trekk, transformerbaserte tekstrepresentasjoner og siamesisk likhetslæring for å etablere et tydelig referansepunkt for videre forbedringer. Kryssmodal justering styrkes deretter gjennom kontrastiv læring i et delt innebyggingsrom, og en multimodal fundamentmodell (LLaVA) tilpasses trafikkdomenet for å forbedre semantisk samspill utover konvensjonelle dual-encoder-arkitekturer. Til slutt analyseres og optimaliseres innebyggingsvalg og strukturert beslutningsnivå-fusjon for å forbedre gjenfinningskvalitet, rangstabilitet, tolkbarhet og beregningsmessig effektivitet. Det foreslåtte rammeverket integrerer multimodal semantisk generering, innebyggingsbasert gjenfinning og eksplisitte konsistensbegrensninger for attributter og bevegelse. Eksperimentell evaluering på CityFlow-NL-benchmarken viser at det komplette systemet oppnår en gjennomsnittlig reciprok rangering (MRR) på opptil 0,80, og leverer ytelse som er konkurransedyktig med sterke tidligere tilnærminger, samtidig som praktisk deployerbarhet ivaretas. Samlet viser avhandlingen at gjenfinning av urbane trafikkscener i stor grad er et systemdesignproblem som krever prinsipiell integrasjon av multimodale representasjoner, bevisst valg av innebygginger og strukturert fusjon som balanserer nøyaktighet, stabilitet og beregningskostnad, og som resulterer i et deployerbart og tolkbart rammeverk for ITS og bredere smarte urbane applikasjoner.

Providing timely and pedagogically useful formative feedback at scale remains a persistent challenge in digital learning environments. Large language models offer a promising approach to automated feedback generation, yet their outputs may not inherently align with established pedagogical quality criteria. This thesis investigates whether embedding explicit assessment and feedback-quality criteria into prompt design influences learners' perceptions of AI-generated formative feedback in introductory programming education. A proof-of-concept system was developed in which participants submitted solutions to short Python programming tasks and evaluated two anonymised feedback versions generated under different prompt conditions: a baseline prompt and a criteria-aware prompt. The criteria-aware prompt incorporated software assessment criteria, feedback quality criteria, a reference implementation, and results from predefined test cases. Both conditions used the same language model configuration, ensuring that observed differences could be attributed to prompt design rather than model variation. Sixteen (16) participants completed this exploratory evaluation. The criteria-aware condition received slightly higher overall ratings and higher ratings on the dimensions of accuracy, specificity, and pedagogical value. However, participants more often preferred the baseline condition, despite rating the criteria-aware feedback higher on several quality dimensions. This suggests that perceived feedback quality and expressed preference do not necessarily align. The results also indicate that the order in which feedback was presented may have influenced participant preferences and that feedback preferences varied between the two programming tasks. Together, these findings suggest that learners' perceptions of AI-generated feedback are influenced not only by feedback quality but also by presentation effects, task characteristics, and learner expectations. The study contributes an empirical illustration of how prompt design influences learners' perceptions of AI-generated formative feedback. The findings suggest that providing more detailed and pedagogically informed feedback does not necessarily lead to stronger learner preference. Effective prompt design may therefore require balancing the amount of information provided with learners' ability to interpret and apply the feedback.

Healthcare shift scheduling in Norwegian municipalities requires simultaneous satisfaction of staffing demand, clinical competence requirements, working-time law, and employee fairness obligations. This thesis investigates whether AI-based scheduling methods can support this problem in a way that accounts for all four dimensions concurrently. A shared scheduling framework is developed around a common domain model, a per-shift legality predicate, and a rolling-horizon legality engine implementing the constraints of labor law. Six method families are implemented and evaluated within this framework: a greedy baseline, mixed-integer programming, tabular reinforcement learning, PPO-based team scheduling, and three families of deep neural network team policies a baseline imitation model, a pairwise candidate scorer, and a contract-goal model that explicitly optimises for contractual hour-target completion. The methods are evaluated across three demand levels representing qualitatively different scheduling problems: single-assignment at demand 1, two-person teams at demand 2, and ten-person teams at demand 10. Results are reported on five hypothesis dimensions covering coverage, overtime, fairness, stakeholder priorities, and optimality gap relative to MIP. All models achieve 100% row-level coverage, but this uniformity is shown to be a weak criterion: the relevant discriminators are horizon completion, workforce utilisation, overtime under the 8-week legal averaging rule, and distributional fairness. Two systematic failure modes are identified across learning-based methods workload concentration and workforce spreading and linked to specific architectural and reward-design choices. The contract-goal DNN model manages both failure modes at demand 10, producing fewer than 220 overtime hours across the full planning horizon, 48-50 completed employee contracts, and a 55-58% average work percentage, competitive with MIP on every critical dimension. At demand 10, the machine learning method family achieves a higher aggregate stakeholder score than exact optimisation (0.72 versus 0.69), providing evidence that learned scheduling policies can surpass exact methods at realistic staffing volumes. The thesis concludes that AI-based healthcare scheduling is viable as decision support, but only under strict constraint-aware design. The strongest results combine learned candidate scoring with explicit legal enforcement rather than treating constraints as soft penalty terms.

The goal of decentralized signal processing is to perform learning tasks from data distributed over a network, where the need for fully decentralized data processing approaches is recognized as a critical issue. Indeed, when data are gathered over a network, sharing local information with a central processor might be not economical or efficient due to several reasons such as size of the network, data volume, energy constraints, or privacy issues. Implementing linear in-network transformations is an application of decentralized signal processing. An important example of in-network transformations is subspace projection, where the aim is to map noisy observations onto a desired signal subspace - a fundamental operation that appears in many problems, including model development, parameter estimation, and other core tasks in signal processing. Recently, graph signal processing has been utilized to implement in-network transformations. The graph signal processing field provides data processing tools such as graph filters for decentralized signal processing. We reviewed existing decentralized methods to implement linear in-network transformations and identified some of the limitations that were not handled previously, which have motivated this Thesis. Most existing methods either consider only special instances of in-network transformations or confine themselves to symmetric networks. In addition, there are some works that restrict themselves to approximate in-network transformations instead of computing them exactly. This thesis proposes several decentralized approaches to implement linear in-network transformations after a finite number of local exchanges. The proposed methods utilize graph signal processing tools, and consider both symmetric and asymmetric networks. In particular, we study the implementation of subspace projection over both symmetric and asymmetric networks. The proposed methods compute the desired projection matrix via graph filters as fast as possible with negligible error. In addition, we consider computing pre-specified linear transformations, and a fast method is proposed via the notion of graph shift operators, which are square matrices that capture the topology of the network. Finally, graph perturbations, which is a common issue in several real applications such as wireless communication networks, are considered in this thesis. First, the negative effect of graph perturbations is analyzed and then an effective approach is proposed to deal with the impact of graph perturbations. Our experimental results demonstrate that the proposed methods outperform current state-of-the-art alternatives.

Supervisory Control and Data Acquisition (SCADA) systems form a critical part of modern critical infrastructure, yet their communication protocols often lack built-in security mechanisms. In particular, the IEC 60870-5-104 (IEC-104) protocol, widely used in power systems, remains vulnerable to cyber threats. Its data transfer mechanisms are susceptible to manipulation and unauthorized access, prompting the need to explore how this risk can be mitigated. To address these challenges, this thesis proposes the design and implementation of an Intrusion Detection System (IDS) aimed at establishing a baseline for identifying anomalous and potentially malicious network activity. Previous research provides valuable and relevant contributions; however, a common limitation is that many studies are older and partly outdated, and some rely solely on general signature-based detection without specific consideration of IEC-104. Using a Design Science Research (DSR) methodology, this thesis develops a baseline framework for IDS detection in IEC-104 SCADA systems. Eight detection rules were designed, each detecting IEC-104 specific packet data that would be considered anomalous and may indicate malicious activity. These rules demonstrate the capability to accurately detect anomalies in both live network traffic and datasets. Finally, using Zeek together with Elasticsearch, Logstash and Kibana, the full IDS test environment is setup with logging and virtualization, while Sigma detection rules are deployed to Kibana using the Lucene query language.

Building Information Modeling (BIM) is increasingly used to support collaboration, information exchange and lifecycle management in the construction industry. As BIM-based environments become more connected through Common Data Environments (CDE), cloud services, external integrations and operational databases, they also become more exposed to cybersecurity risks. This is especially important when the information contains critical infrastructure-related data, such as technical room information, building layouts, infrastructure details and operational information. This thesis examines how structured threat modeling can be used to identify and prioritize cybersecurity risks in a BIM-based environment, where project information is transformed and reused for operational purposes. The study is based on a conceptual process architecture inspired by case material from Avinor. The system is modeled using the Microsoft Threat Modeling Tool and analyzed through the STRIDE threat categorization framework. Selected threat scenarios are then prioritized using DREAD risk assessment model and mapped to relevant NIST SP 800-53 security controls. The analysis shows that BIM-based environments should be understood as connected information systems rather than only as design or collaboration tools. The main risks identified are related to information disclosure, elevation of privilege, spoofing, tampering and weak traceability. Three threat scenarios are analyzed in detail: unrestricted access to the CDE, excessive access to sensitive technical room information and insufficient filtering during the transformation of project information into operational information. The results show that cybersecurity risks may arise not only within individual components, but also through data flows, trust boundaries and lifecycle transitions. The thesis concludes that structured threat modeling can support a more systematic understanding of security risks in BIM-based environments. It also shows that mitigation should focus on strong authentication, role-based access control, least privilege, information classification, validation, filtering, logging and traceability.

Cyber ranges are increasingly used in cybersecurity education because they provide realistic and isolated environments for practical training. However, existing research shows that their educational value depends not only on technical realism, but also on how training activities are structured, executed, and evaluated. This thesis addresses that challenge by designing, implementing, and evaluating a structured cyber range training solution for SOC exercises in higher education. The proposed solution was developed using a Design Science Research approach and combines CyberRangeCZ, OpenStack, and the Malcolm monitoring platform into a training environment. The implemented exercise is based on a web application compromise and guides students through a staged investigation workflow involving scenario briefing, log analysis, guided tasks, and a final incident report. To support systematic assessment, the thesis introduces a multi layer evaluation framework that combines automated task scoring with rubric evaluation of the incident report. This provides a structured approach for assessing both task performance and analytical reasoning. The solution was evaluated through participant testing and a post training survey. The results indicate that the implemented cyber range exercise can support SOC training and assessment by combining realistic monitoring data, structured investigation tasks, and reflective reporting. Participants generally perceived the workflow as understandable and the training as usable, realistic, and educational. They also reported positive learning outcomes related to log and alert analysis, identification of indicators of compromise, and understanding of SOC investigation workflows. At the same time, the findings show that some usability challenges remained, particularly in relation to extracting relevant information efficiently from Malcolm and understanding certain task formulations. Overall, the findings suggest that open source cyber range environments can provide a viable basis for SOC training in higher education when they are combined with structured scenario design, monitoring, and systematic evaluation. The work contributes both a concrete implementation and a reusable design perspective for integrating training workflow and assessment in cyber range cybersecurity education.

English: Maritime organisations increasingly rely on cloud-connected software for operational, administrative, compliance, and ship-to-shore workflows. This thesis assesses whether selected maritime cyber risk frameworks provide traceable threat-level coverage for such systems, using a cloud-integrated maritime information platform as a case study. The study combines a structured literature review, case-based data flow modelling, STRIDE threat modelling, maritime framework analysis, and qualitative validation with the industry partner. The literature review produced 24 included papers and was used to derive a consolidated STRIDE catalogue of 101 threats relevant to maritime and cloud software contexts. The case platform was modelled through three data flow diagrams covering cloud access, internal application modules, and ship-side/offline operation. The resulting threat catalogue was applied to the model and mapped against six maritime cyber risk sources relevant to the Norwegian context. The analysis shows that the framework set provides a governance and resilience baseline while explicit threat-level traceability remains incomplete for a cloud-integrated maritime software system. Across the 101 threats, 76 had explicit coverage in at least one selected source, 24 were only implicitly covered, and one had no traceable coverage. IACS UR E27 provided the strongest explicit technical coverage. The weakest areas concerned cloud-native and AI-related threats, Software as a Service and external dependency outages, and application-level security threats that are not directly expressed in several maritime sources. To address this gap, the thesis proposes a software development life cycle extension that connects DFD-based STRIDE analysis, threat registers, framework coverage mapping, and Elevation of Privilege workshops. The workshop and selected industry partner validation supported the practical relevance of the approach, while full operational integration remains future work. The thesis concludes that maritime cyber risk frameworks should be supplemented with system-specific threat modelling to achieve traceable coverage of cloud-integrated maritime software threats. Norsk: Maritime organisasjoner er i økende grad avhengige av skybasert programvare for operasjonelle, administrative og regelverksrelaterte arbeidsprosesser, samt samhandling mellom skip og land. Denne oppgaven undersøker om utvalgte rammeverk for maritim cyberrisiko gir sporbar dekning av trusler mot slike systemer, med en skyintegrert maritim informasjonsplattform som casestudie. Studien kombinerer et strukturert litteraturstudie, casebasert modellering av dataflyter, STRIDE-trusselmodellering, analyse av maritime rammeverk og kvalitativ validering i samarbeid med en industripartner. Litteraturgjennomgangen resulterte i 24 inkluderte artikler og dannet grunnlaget for en samlet STRIDE-katalog med 101 trusler relevante for maritime og skybaserte programvaresystemer. Casestudien ble modellert ved hjelp av tre dataflytdiagrammer som beskriver skytilgang, interne applikasjonsmoduler og drift på skipssiden. Trusselkatalogen ble deretter anvendt på modellen og kartlagt mot seks kilder for maritim cyberrisiko som er relevante i norsk kontekst. Analysen viser at de utvalgte rammeverkene gir et godt grunnlag for styring av cybersikkerhet og organisatorisk motstandsdyktighet, men at sporbarheten mellom konkrete trusler og rammeverkskrav fortsatt er ufullstendig for skyintegrerte maritime programvaresystemer. Av de 101 identifiserte truslene hadde 76 eksplisitt dekning i minst én av kildene, 24 var kun implisitt dekket, og én trussel manglet sporbar dekning. IACS UR E27 ga den sterkeste eksplisitte tekniske dekningen. De største svakhetene gjaldt sky- og KI-relaterte trusler, driftsavbrudd i SaaS-tjenester og eksterne avhengigheter, samt sikkerhetstrusler på applikasjonsnivå som ikke er tydelig omtalt i flere av de maritime kildene. For å håndtere denne utfordringen foreslår oppgaven en utvidelse av industripartnerens livssyklus for programvareutvikling (SDLC) som kobler sammen DFD-basert STRIDE-analyse, trusselregistre, kartlegging av rammeverksdekning og Elevation of Privilege-workshops. Resultatene fra workshopen og valideringen med industripartneren støtter den praktiske relevansen av tilnærmingen, mens full operasjonell integrering er overlatt til framtidig arbeid. Oppgaven konkluderer med at maritime rammeverk for cyberrisiko bør suppleres med systemspesifikk trusselmodellering for å oppnå sporbar dekning av trusler mot skyintegrert maritim programvare.

The rapid expansion of Internet of Things (IoT) devices has made IoT networks a major target for cyberattacks, while the resource-constrained nature of IoT hardware limits the security mechanisms that can be deployed. This thesis addresses two complementary detection problems: anomaly detection for IoT sensor data and flow-based intrusion detection. A physical smart home testbed was designed and deployed, including a Raspberry Pi gateway, five ESP8266 + DHT22 temperature and humidity sensors, a PySense sensor, and three Kali Linux virtual machines acting as attackers. A custom lightweight network flow extractor was implemented directly on the Raspberry Pi, producing a labeled dataset of 75 flow features across 54 attack sub-types in 10 categories, including False Data Injection and Backdoor categories absent from existing public datasets such as NSL-KDD, UNSW-NB15, and CIT-IoT-2023. For sensor anomaly detection, an LSTM-autoencoder was trained exclusively on benign sensor readings. Evaluated against hardware faults, direct data injection, and MITM replay attacks. The model achieved 98.70\% accuracy, 96.25% F1 score, and 99.40% AUC-ROC (with replay data excluded), confirming that it is effective for distinguishing sensor anomalies but fundamentally unable to detect replay attacks whose data is indistinguishable from normal traffic. For intrusion detection, XGBoost was established as a supervised baseline, achieving 95.8% accuracy and 94.6% macro F1 on 54-class classification of the custom dataset. Model-Agnostic Meta-Learning (MAML) and Prototypical Networks (ProtoNet) were evaluated under a 5-way classification of the custom dataset on 10 novel attack classes unseen during meta-learning. ProtoNet achieved 96.88% at 5-shot and 94.30% at 1-shot, outperforming MAML (95.93% and 91.62%). These results demonstrate that few-shot learning is a viable approach for intrusion detection in IoT networks, enabling rapid adaptation to novel attack categories without retraining.

Denne masteroppgaven undersøker hvordan Web3-teknologier og hybrid homomorfisk kryptering kan støtte sikker og personvernbevarende deling av data i IoT miljøer. Den foreslåtte arkitekturen kombinerer IPFS for lagring utenfor blokkjeden, IOTA for desentralisert koordinering av metadata, og hybride homomorfiske krypteringsskjemaer for kryptert aggregering av sensordata. En modulær eksperimentell prototype ble implementert og evaluert ved hjelp av en klartekst-baseline, PASTA-basert kryptering, HERA-basert kryptering og to strategier for datagranularitet: finmasket publisering og hybrid batching. Evalueringen viser at IPFS og IOTA kan utgjøre et gjennomførbart desentralisert grunnlag for deling av IoT-data ved å skille datalagring fra koordinering. IPFS tilbød stabil lagring utenfor blokkjeden for både klartekst og krypterte nyttelaster, mens IOTA muliggjorde sporbar publisering av metadata og innholdsidentifikatorer, men med høyere og mer variabel kommunikasjonsforsinkelser. Resultatene viser videre at hybrid homomorfisk kryptering forbedrer konfidensialiteten ved å tillate aggregering uten å eksponere klartekstverdier, men introduserer betydelig beregningsmessig overhead, særlig under kryptert aggregering. Hybrid batching reduserte gjentatte publiserings- og prosesseringskostnader sammenlignet med finmasket publisering, men introduserte også avveininger knyttet til responstid og datagranularitet. Masteroppgaven bidrar med en eksperimentell Web3-basert arkitektur for personvernbevarende deling av IoT-data, en komparativ evaluering av klartekst- og krypterte konfigurasjoner ved bruk av PASTA og HERA, samt en analyse av avveiningene mellom personvern, ytelse, skalerbarhet og desentralisering. Samlet sett indikerer funnene at desentralisert deling av IoT-data med hybrid homomorfisk kryptering er teknisk gjennomførbart, men at praktisk implementering avhenger av tilgjengelig arbeidsmengde, maskinvarekapasitet, krav til nettverksforsinkelser og personvernbehov.

Cyber ranges are increasingly used in cybersecurity education because they provide realistic and isolated environments for practical training. However, existing research shows that their educational value depends not only on technical realism, but also on how training activities are structured, executed, and evaluated. This thesis addresses that challenge by designing, implementing, and evaluating a structured cyber range training solution for SOC exercises in higher education. The proposed solution was developed using a Design Science Research approach and combines CyberRangeCZ, OpenStack, and the Malcolm monitoring platform into a training environment. The implemented exercise is based on a web application compromise and guides students through a staged investigation workflow involving scenario briefing, log analysis, guided tasks, and a final incident report. To support systematic assessment, the thesis introduces a multi layer evaluation framework that combines automated task scoring with rubric evaluation of the incident report. This provides a structured approach for assessing both task performance and analytical reasoning. The solution was evaluated through participant testing and a post training survey. The results indicate that the implemented cyber range exercise can support SOC training and assessment by combining realistic monitoring data, structured investigation tasks, and reflective reporting. Participants generally perceived the workflow as understandable and the training as usable, realistic, and educational. They also reported positive learning outcomes related to log and alert analysis, identification of indicators of compromise, and understanding of SOC investigation workflows. At the same time, the findings show that some usability challenges remained, particularly in relation to extracting relevant information efficiently from Malcolm and understanding certain task formulations. Overall, the findings suggest that open source cyber range environments can provide a viable basis for SOC training in higher education when they are combined with structured scenario design, monitoring, and systematic evaluation. The work contributes both a concrete implementation and a reusable design perspective for integrating training workflow and assessment in cyber range cybersecurity education.

Mobile communication has evolved significantly since the introduction of the zero generation mobile system in 1946, until the present fifth generation mobile network. Capacity, coverage and security improvements have been a central part of the mobile network evolution. In this context, this thesis investigates the use of unmanned aerial vehicles (UAVs) as an extension to 5G networks to enhance coverage and system capacity while maintaining strong security guarantees. To support UAV assisted communication, a set of protocols is proposed based on the 5G architecture. These include authentication protocols, data transfer mechanism, and handover procedures. The proposed solutions are evaluated through both performance and security analysis. The performance analysis demonstrates that integrating UAVs into existing 5G networks can effectively extend coverage and increase capacity, while preserving the user experience across multiple applications. The security analysis shows that these improvements can be achieved without compromising security. The proposed design ensures mutual authentication between user equipment (UE), UAVs, and the base station (gNB), while protecting sensitive communication. Furthermore, the protocols are designed to align closely with existing 5G standards, enabling integration with minimal modifications. Overall, the results indicate that UAV assisted communication is an enhancement to current 5G systems. The thesis also outlines directions for future work to further evaluate and refine the proposed solution.

Generative Artificial Intelligence (GAI) has become increasingly integrated into real estate visualization workflows. This approach allows for the quick creation of photorealistic exterior and interior designs. However, current AI-supported image editing approaches offer limited support for an interactive, precise revision workflow. This implies that minor adjustments often affect not-intended areas of the image, leading to inconsistent lighting, texturing, and scene structure. As a result, professionals frequently rely on multiple tools to complete revision tasks, reducing process efficiency. This thesis explores a possibility of user-controlled interactive revision of AI-created exterior and interior designs while maintaining visual consistency. The research followed a Design Science Research (DSR) methodology combined with principles of User-Centered Design (UCD). A literature review, tool testing, and expert discussions were executed to identify workflow limitations with existing design tools, and to define system requirements. Based on the results, a prototype system, ImaReview, was designed and implemented with the goal to support interactive image revisions through brush region selection, interactive object placement, and text instructions. The system was evaluated through internal testing and user studies, with thematic analysis and System Usability Scale (SUS) assessments. The outcomes show that users perceived the prototype as intuitive and useful for interactive design workflows, specifically due to its localized editing capabilities and simplified interaction process. The findings show the potential of interactive AI-assisted revision workflows for professional real estate visualization and point out the importance of controllability and structural consistency in future generative AI editing systems.

Hyperspectral image classification plays an important role in remote sensing applications such as land-cover mapping and environmental monitoring. However, many existing classification approaches rely on deep learning models evaluated under pixel-level random train-test splits, which may introduce spatial information leakage and lead to overly optimistic performance estimates. At the same time, interpretable and data-efficient alternatives remain relatively underexplored for hyperspectral signal classification. This thesis presents a patch-level hyperspectral classification framework based on the Tsetlin Machine together with a structured Boolean encoding strategy and a clause-level explainability analysis pipeline. Instead of operating directly on high-dimensional spectral cubes, the proposed method extracts statistical spectral descriptors from image patches using band-wise mean and standard deviation features. These descriptors are then transformed into Boolean literals through a multi-stage encoding framework combining quantile-based thresholds, local spectral structure detection, and derivative-based tests. This process produces an informative binary representation suitable for clause-based learning. Multiple binarization strategies were evaluated to analyze their impact on classification performance. The results show that a three-level discretization approach achieved the best performance, while PCA-based dimensionality reduction reduced both classification accuracy and feature interpretability. In addition, a comparison between Vanilla and Coalesced Tsetlin Machines demonstrated that clause sharing improves classification performance while preserving interpretability. To improve the explainability of model behaviour, Random Forest importance ranking was used to identify the most influential clauses contributing to classification decisions. The analysis revealed that only a small subset of clauses dominates the decision process and captures meaningful reflectance magnitude and spectral variability patterns. Furthermore, discrimination analysis showed that the spectral interval between bands 10 and 75, especially 35 and 60, plays a key role in separating vegetation classes in the Salinas dataset. Overall, the results demonstrate that combining structured Boolean encoding with clause-based learning provides an effective, interpretable, and computationally efficient framework for patch-level hyperspectral image classification under leakage-free evaluation settings.

This thesis presents an automated processing pipeline for Norwegian veterinary insurance claims, developed inside the privacy and operational boundary of a Norwegian insurance company. Manual claim handling is a time-consuming bottleneck during high-volume peri- ods, and the work documented here is the company’s pilot effort at building an end-to-end pipeline that can decide the cases it is capable of deciding and route the rest to existing case-handler tooling. The pipeline is implemented as a LangGraph state machine of 13 nodes that combines three complementary extraction strategies (regular expressions, local GLiNER named-entity recognition and custom parsers), a Retrieval-Augmented Generation step over indexed insurance policy terms, a confidence gate that routes uncertain claims to manual review and a template-driven output layer that produces payment instructions, customer-facing emails and internal case notes from operator-approved Norwegian templates. The assessment LLM (gpt-4.1-mini) runs in the company’s Azure-OpenAI tenancy on payloads from which the customer’s name and PII fields have been stripped on the local machine before any cloud call, because clearance to process customer personal data on the cloud endpoint could not be established within the project’s time budget. Across a manual outcome log of 73 test claims, the pipeline produced a correct fully- automated decision on 60% of cases, surfaced 34% as extraction failures that an operator marked for manual handling and produced a wrong denial on 5% of cases, with no wrong approvals. Three of the four wrong denials cluster in a single clinic template, and in all four the extraction step misread the journal and produced a wrong diagnosis that the assessor then evaluated. The absence of wrong approvals is the load-bearing observation against the company’s financial-leakage criterion (approximately 3.5% for ordinary handler workflows, per a short post-demonstration interview with the company’s contact); the wrong-denial cases generate complaint and rework cost rather than incorrect payouts. The interview also confirmed that the system addresses the priorities the company cares about (reducing claim- handler workload at high volume, lowering operating cost and shortening customer response times). The principal architectural compromise this thesis was forced to make on privacy grounds, the cloud boundary at the assessment step, would be resolved by migrating the LLM to a company-internal endpoint, which is identified as the highest-priority technical extension.

This thesis investigates whether maintenance cost patterns extracted from housing association annual reports can improve property insurance risk prediction. An automated pipeline converts semi-structured annual-report PDFs into structured financial data using document layout analysis and Large Language Models (LLMs), processing 26 580 documents and reaching 91.6% agreement on net income (within ±1%) against an independent commercial database. From these extractions we assemble a panel of approximately 1 400 Norwegian housing associations whose annual reports span 2010–2024, linking public financial data from the Brønnøysund Register Centre with property characteristics from the national cadastre, matched to Tryg’s internal claim records over a 2015–2023 validation window. We compare a graduated hierarchy of ten model architectures across five families (Generalised Linear Models (GLMs) and Generalised Additive Models (GAMs), glass-box models, gradient-boosted ensembles, neural architectures, and a logic-based Tsetlin machine) on both predictive accuracy and interpretability. Engineered temporal features (multi-year maintenance slopes, volatility, and spikes) test the hypothesis that deferred maintenance elevates future claims. On the public premium proxy, adding the maintenance features to a Tweedie GLM changes deviance only marginally (−0.9%, reversing to +3.7% once temporal lags are added). The substantial gains come from the model class, not the feature augmentation: the best ensemble cuts Tweedie deviance by 36.7% over the GLM on the control features (Gini 0.743), yet no architecture extracts net premium value from the maintenance block. Because the premium embeds the incumbent insurer’s own rating, the decisive test is against actual claims, where maintenance features carry a limited signal: they predict claim frequency (Gini 0.621) but not per-claim severity (Gini in [−0.07, +0.08]), so maintenance-based pricing should target the frequency loading. A forward-horizon analysis shows maintenance features track trailing claims history closely (one-year Gini gap 0.005), making them a practical near substitute for claims history when underwriting established associations that lack a claims record with the insurer.

Large language models have gone from small next-token predictors to highly sophisticated systems. As these systems have grown, a key challenge has been to make sure they behave in line with human values and intentions, a goal often referred to as alignment. In recent years these models have increasingly been evaluated for alignment and safety, but many evaluation approaches emphasize final responses. Yet the final response is only the end of what a model produces, and by itself it reveals little about the process that produced it. One way to make that reasoning visible is chain-of-thought, where the model produces intermediate reasoning steps in readable text before its final answer. This began as a prompting technique, but more recently a class of models known as reasoning models has emerged, producing such reasoning natively in a dedicated part of the output. This thesis studies this chain-of-thought reasoning, both prompted and native, and compares it across models and task formats. The study treats the reasoning not as direct evidence of internal decision-making, but as observable model output that can be analyzed alongside the final answer. Three models with different origins are evaluated, Qwen3-8B, NorMistral-11B-thinking, and Llama 3.1-8B-Instruct. They are tested on two datasets, MoralChoice, a set of 1,367 binary ethical dilemma scenarios with labeled moral violations, and Do-Not-Answer, a set of 939 harmful prompts. The evaluation combines final-output metrics with analysis of the reasoning and an LLM-as-a-judge setup that uses two independent judges, checked against the author's own annotations on a smaller set of traces. The results show that the two datasets measure different kinds of behavior and do not combine into a single picture of which model is most aligned. On the ethical scenarios where one action is clearly the better choice, all three models almost always avoid the harmful one, with differences too small to separate them. On the ambiguous scenarios, where neither action is clearly better, the models disagree more often, which might reflect different moral preferences rather than one model being more harmful. On the harmful prompts the differences are clearer, with Llama 3.1 classified as following harmful instructions most often and NorMistral least often. In the judged traces, the stated reasoning and the final answer generally matched on the ethical scenarios, where a model's reasoning argued for the action it then chose. On the harmful prompts the picture was less consistent. In some traces a model noted that a request was sensitive or that it should be careful, and then answered it anyway. Overall, the reasoning traces are better treated as a supplementary signal for studying misalignment than as a standalone way to identify it.

The applicability of machine learning (ML) models for deployment on resource-constrained edge devices has become increasingly relevant with the growth of the Internet of Things (IoT) and embedded systems. Anomaly detection is particularly important in such environments, where reliable and efficient models are required to identify abnormal behavior in real time. However, strict limitations in memory size, computational capacity, and energy availability make the deployment of ML models on microcontroller (MCU)-based platforms challenging. This thesis investigates lightweight one-class classification (OCC) models for edge-based anomaly detection, with a focus on comparing three Tsetlin Machine (TM)-based approaches and a neural network (NN)-based autoencoder approach. TMs represent a promising logic-based learning paradigm, offering interpretable rule-based representations and potentially efficient inference for resource-constrained systems. Both model types are implemented and deployed on an ESP32-C6 MCU to evaluate their behavior under realistic hardware constraints. The models are evaluated on multiple datasets, including industrial supervisory control and data acquisition (SCADA) data, electrocardiogram (ECG) signals, and occupancy data, where the evaluation considers predictive performance, memory usage, inference time, and energy consumption. The results show that model suitability is highly dependent on the application context and the constraints of the deployment environment. While TM-based models offer advantages related to interpretability and compact logical representations, they currently lack the optimization maturity of TensorFlow-based NN deployments. As a result, the TensorFlow Lite for Microcontrollers (TFLM)-based models achieve stronger deployment efficiency in several experiments, while the TM-based approaches demonstrate potential for further improvement through dedicated embedded optimization. Overall, the results reveal a multidimensional trade-off between predictive performance and resource efficiency. This emphasizes the importance of context-aware model selection when designing embedded anomaly detection systems.

The continuous growth in connected devices, Internet of Things (IoT) systems, and bandwidth-intensive applications has significantly increased the demand for wireless network capacity and performance. While modern cellular technologies such as Fourth Generation (4G) and Fifth Generation (5G) mobile systems provide improved spectral efficiency, low latency, and support for large numbers of users, expanding cellular infrastructure remains costly due to the limited availability of licensed spectrum. At the same time, WiFi technologies operating in the unlicensed spectrum have evolved to provide high-capacity wireless connectivity through dense and cost-efficient deployments. This has ignited interest in Carrier WiFi solutions, through which WiFi networks are integrated with mobile operator infrastructure to offload traffic from cellular networks. However, telecommunication infrastructures and Carrier WiFi solutions have traditionally been dominated by proprietary and closed ecosystems, limiting flexibility, interoperability, and accessibility for experimentation and innovation. This thesis showcases the design, implementation, and evaluation of an open-source Carrier WiFi architecture enabling Subscriber Identity Module (SIM)-based authentication for WiFi access. Our solution is primarily built based on Magma and integrates WiFi infrastructure with existing mobile Core Network (CN) components, supporting seamless interoperability between WiFi and cellular networks. By leveraging authentication mechanisms such as Extensible Authentication Protocol (EAP)-AKA and Passpoint, the system enables seamless onboarding and authentication of Commercial Off-The-Shelf (COTS) User Equipment (UE) using SIM credentials only, without requiring manual user configuration. The work further investigates differentiated Quality of Service (QoS) policies and subscriber-specific traffic handling within the Carrier WiFi network. Extensive experimental evaluation was conducted using multiple commercial UEs to assess authentication functionality, global end-to-end connectivity, policy enforcement, network performance, and inter-Radio Access Technology (RAT) handoff between mobile and WiFi networks. Through real-life experiments, we demonstrate the successful implementation of a open-source Carrier WiFi architecture supporting SIM-based authentication, differentiated service handling, and seamless access across WiFi and cellular networks while maintaining a satisfactory user experience.

Large Language Models (LLMs) are increasingly explored for public-sector tasks in which outputs must be reliable, auditable, and useful within domain-specific workflows. This the- sis explores the potential of, and develops a prototype for, a multi-dimensional benchmark- ing pipeline for evaluating locally deployed LLMs across summarization quality, semantic similarity, faithfulness-oriented scoring, demographic bias, medical bias classification, parse reliability, latency, and domain-specific quality assessment. The prototype is applied to locally deployed open-weight models in the 2.7B–20B range. A proprietary GPT-5-nano baseline, provided by the Norwegian Institute of Public Health (FHI), is included where comparable outputs are available. The central case study is AMSTAR-2 systematic review appraisal: 25 systematic review articles on screen time and child/adolescent development are evaluated against FHI expert reference ratings across all 16 AMSTAR-2 items and the overall confidence rating. The results show that no single metric provides an adequate picture of model suitability. In biomedical summarization, lexical and semantic metrics can reward vague summaries that contain little substantive information. In the bias experiments, occupation-related generation reveals systematic gender associations, while medical bias classification remains difficult. A re-scored run with a corrected prompt and an explicit None option shows persistent default- label collapse, with models often over-predicting one class. In the corrected AMSTAR-2 experiment, no model exceeds an article-independent majority-class accuracy baseline, so item accuracy is interpreted through chance-corrected agreement: gemma3:4b, mistral:7b, and gpt-oss:20b reach item-level agreement with the FHI expert ratings comparable to the available GPT-5-nano baseline, whereas GPT-5-nano performs best on overall confidence ratings. As the GPT-5-nano outputs were produced under different prompting and on a smaller article set, this is a bounded comparison rather than a controlled ranking. A key methodological finding is that structured prompts can introduce validity artefacts. In the initial AMSTAR-2 run, several models copied a concrete JSON example from the prompt rather than assessing the article. Response-variance analysis detected this template-copying behaviour, and a corrected placeholder-based prompt changed the model rankings. The thesis therefore argues that LLM evaluation for public-sector use should be task-specific, multi-dimensional, and explicitly validated for prompt sensitivity, structured-output reliabil- ity, and response variance. The evaluated local models appear promising as decision-support tools, but not reliable enough for autonomous quality appraisal without expert review.

Open-source implementation represents a promising tendency for private 5G network development and deployment. Representatively, several open-source 5GC projects have been developed in recent years for research and development purposes, making it possible to study, test, and evaluate real 5G network functions in flexible experimental environments. Based on free5GC, this thesis project designs, implements, and evaluates an end-to-end 5G system which supports both 3GPP access and untrusted non-3GPP access. The designed system contains one common network architecture and is implemented in two operation modes. Operation Mode-V is implemented as a virtual testbed using free5GC, UERANSIM, and N3IWUE, while Operation Mode-H is implemented as a hardware-based testbed using free5GC, OAI RAN, OpenWiFi, SDR hardware, COTS UEs, and N3IWUE. Through one working system which supports both operation modes, we showcase the operability and interoperability of both 5G and WiFi networks through one common core network platform. Furthermore, extensive experiments have been performed to evaluate connectivity, throughput, latency, packet loss, and QoE. The results show that both 3GPP access and non-3GPP access were successfully established through the common 5GC. In the final system, the VM hosting the free5GC CN was allocated 8 CPU cores and 8192 MB RAM, the highest measured DL throughput at the free5GC VM reached approximately 127 Mbps (Ookla), while the UE-side iperf3 measurement showed an average of approximately 117 Mbps for a single connected 5G UE with 0\% packet loss. In addition, the QoE evaluation demonstrated successful 2160p (4K) video streaming without buffering, confirming the capability of the implemented system to support bandwidth demanding user applications.

This thesis investigates the application of the Convolutional Tsetlin Machine (CTM) to medical image classification in terms of classification performance and clause-based decision transparency. The investigation was conducted using three datasets from the MedMNIST benchmark: BreastMNIST, PneumoniaMNIST, and BloodMNIST. For each dataset, five CTM specialists were trained, each using a version of the dataset booleanized by a distinct technique. The booleanization techniques employed include Canny Edge Detection, Adaptive Gaussian Thresholding, Adaptive Mean Thresholding, Otsu's Thresholding, and Thermometer Encoding. Furthermore, composite combinations, in which multiple specialists contribute to a joint classification decision, were evaluated alongside the individual specialists. Classification performance was evaluated using several performance metrics such as accuracy, area under the curve, and per-class recall. To evaluate clause-based decision transparency, statistical analysis of clause length distributions was performed along with the analysis of the spatial distributions of clause activations. Statistical analysis comprised boxplots, Mann-Whitney U tests with rank-biserial correlation effect sizes to compare clause length distributions across specialists for a given class and polarity, as well as between positive and negative polarities for a given class within each specialist. The analysis of spatial distributions was performed using the proposed visualization methodology, comprising local and global class representations of positive polarity and negative polarity clause activations for each trained CTM specialist. Experimental results demonstrated that the Thermometer specialist generally achieved the strongest classification performance across all three datasets. Composite combinations formed from subsets of specialists improved upon the performance of individual specialists in several configurations. Regarding clause structure, the Thermometer specialist generally had longer clause representations, while within-specialist polarity differences in clause length were negligible in practical terms, despite reaching statistical significance in some cases. The spatial activations mostly reflected the booleanized images, with negative literals focusing on the black pixels and positive literals focusing on the white pixels, with each specialist producing distinct activation patterns consistent with its booleanization strategy. The competitive performance demonstrated by the CTM specialists and composites along with the demonstrated clause-based decision transparency suggest that the CTM is a promising approach for medical image classification tasks, offering both strong performance and inspectable decision-making. Future work could explore further optimization of the CTM architecture and training procedures, as well as the application of the proposed visualization methodology to other domains and datasets. Most importantly, future work should evaluate the interpretability of the clause-based representations with domain experts in clinical settings.

The industry experiences a rapidly growing tendency for developing multiple solutions for LPWAN and IoT networks, with DECT NR+ (DECT-2020 NR) being a recent addition. Existing evaluations of DECT NR+ mesh capability have been largely confined to simulation and single-hop link measurements, leaving the multi-hop transport of image-sized payloads on real-world hardware largely unexplored. This thesis designs, implements, and evaluates a DECT NR+ mesh network capable of transmitting real-time images over multiple hops on COTS hardware. The system is built on the Nordic Semiconductor nRF9151-DK and consists of three component roles: PT for leaf nodes, FT for the sink, and a paired FT/PT node as a relay, working together with a full-stack application that displays image data and reports per-image metrics through a web interface. Across indoor, semi-outdoor, and pure-outdoor experiments, the implemented network reaches 120 m on a single indoor hop, extending up to 180 m once a second hop continues outdoors, and 400 m on a single hop extending up to 660 m with two hops in a pure-outdoor environment. In an outdoor environment, single-hop links achieve greater than 99% Packet Delivery Ratio (PDR) with an estimated median end-to-end latency as low as 110–150 ms, while two- and three-hop configurations drop to roughly 69–72% PDR with estimated median latencies between 600 ms and over 1300 ms. Our results demonstrate the feasibility and operability of a developed DECT NR+ mesh network capable of transmitting images over multiple hops based on COTS hardware. In the meantime, our study also identifies the specific constraints of the adopted hardware and reveals the need for an ETSI standards-compliant single-device relay.

Massive Internet of Things (mIoT) networks support many devices that transmit short and sporadic uplink packets. For this traffic type, conventional connection-based access can introduce signalling overhead that is heavy compared with the payload, while grant-free access can increase collision pressure when many devices transmit simultaneously in the same frame. This thesis develops and evaluates a self-contained frame-based Medium Access Control (MAC)-level Python simulator with a Web-Based User Interface (Web-UI) for smalldata uplink access in clustered Non-Orthogonal Multiple Access (NOMA) IoT networks. The work is associated with Fifth Generation (5G) New Radio (NR) Small Data Transmission (SDT), grant-free access, hashing-based slot selection, and Reinforcement Learning (RL)- based access control, but the simulator is intentionally SDT-enabled and NOMA-enabled rather than a complete 3GPP NR or physical-layer implementation. The simulator models logical device clusters, Bernoulli packet arrivals, finite frame slots, per-device queues, access control, uniform and hash-based slot selection, feedback, retransmission, deadline handling, packet drops, MAC-level performance metrics, and a probabilistic detection abstraction based on Table-II in a reference paper. A user interface enables fixed, random, tabular Q-learning, and Deep Q-network (DQN) policies to select per-cluster access probabilities. The Web-UI provides a practical front-end for launching experiments, replaying saved runs, comparing results, and inspecting generated artefacts, supporting an artefact-based evaluation workflow that covers detector validation, access-method comparison, traffic and network-scaling sensitivity, static access tuning, hash and hybrid-hash sensitivity, learning-controller diagnostics, reward-design diagnostics, and MAC-sensitivity studies. The detector validation shows close agreement between configured and empirical Table-II detection probabilities. In the access-method comparison, the fully hash-based controlled scheme, Scheme C, achieves the highest successful packets per frame and Packet Success Ratio (PSR) among the evaluated Scheme A, Scheme B, and Scheme C configurations across the tested network scenarios and offered loads. However, the results also show important tradeoffs. Without-Access-Control (WAC) operation remains competitive at low load, Scheme A often provides higher Jain fairness under contention, and high-load operation remains overloaded even under Scheme C. The static access-probability sweep further shows that static access tuning matters because the best shared access probability depends on the offered load. The matched learning-controller comparison shows that DQN improves over tabular Q-learning in throughput and PSR, however this comes at significant fairness cost. Q-learning is generally fairer, while tuned static access remains the stronger throughput baseline. What is contributed in this thesis is a self-contained simulator with a Web-UI and an evaluation workflow for studying SDT/NOMA-enabled MAC-level access mechanisms under common assumptions, with the results being interpretable within a simulation and NOMA physical-layer abstracted boundary.

This thesis investigates how autonomous mobile robots in warehouse environments can provide more useful information after a safety-related stop. The work is motivated by the practical problem that warehouse operators often receive little diagnostic information when an autonomous robot stops. By adding semantic post-stop assessment, the system aims to support faster and more informed decision-making, reduce unnecessary manual inspection, and improve the operational usefulness of autonomous pallet trucks in dynamic warehouse environments.

This thesis proposes and evaluates a Tsetlin Machine (TM) autoencoder for end-to-end wireless communication. In autoencoder-based communication systems, the encoder and decoder are trained jointly to learn a representation adapted to the channel. Since the Tsetlin Machine does not support gradient-based training, a decoder feedback mechanism is developed to generate training targets for the encoder, allowing the system to be trained without backpropagation. Two research questions are investigated. First, we investigate whether the training of the TM autoencoder can be improved through changes to the system. Second, we investigate whether the TM autoencoder can handle a two user scenario, where two users transmit simultaneously over the same channel. The system is evaluated on the additive white gaussian noise (AWGN) channel and the Rayleigh fading channel, and compared against a deep neural network (DNN) autoencoder. The results show that the TM autoencoder matches the performance of the DNN autoencoder on the AWGN channel and is competitive on the Rayleigh fading channel. It also performs well at low $E_b/N_0$ (signal to noise ratio) for the two user scenario. The main limitation is a reduced rate of improvement at high $E_b/N_0$ ratio, which appears consistently across all scenarios. To the best of our knowledge, this is the first demonstration of both a Tsetlin Machine in an end-to-end communication system and a TM encoder decoder training.

Few-shot learning, continual learning, and test-time adaptation are each well-studied individually, but no architecture currently addresses them jointly. The combined problem asks for a system that meta-trains offline and then specializes itself, at inference, to a small stream of examples drawn from classes it has never seen, without modifying its weights and without the second-order gradient computations characteristic of MAML-style meta-learning. This thesis proposes CLEM: a Cyclic Latent Embedding Meta-learner that confines all task-specific adaptation to a single latent vector. CLEM combines a dual-stream Vision-Transformer encoder–decoder, a classifier head, and a small weight-shared Learning block that updates the latent vector at every step of a support sequence. Training optimizes a composite of image-reconstruction, latent-reconstruction, and a clipped learning-improvement loss comparing the classifier’s cross-entropy on a held-out test image with the initial vs. post-sequence latent vector. The clipping bound, pinned at the random-chance loss log K, revents the optimizer from gaming the improvement signal by degrading its zero-shot baseline below chance. The architecture is evaluated on CIFAR-100 under a disjoint-class protocol that tests on classes never seen during meta-training. A grid of 4 × 5 × 3 = 60 hyperparameter configurations over latent dimension d, total class count N , and sequence multiplier m, with 20 independent runs per cell, produces 1200 training runs. The no-learn baseline tracks random chance at every N , as the loss design enforces. The principal finding is a monotone scaling trend: the mean relative improvement of the Learning block rises from −7.77% at N =10 to +13.04% at N =100 (Spearman r = +0.165, p < 10−8). The configuration (d=128, N =100, m=4) reaches statistical significance at p = 7×10−4 with a large effect size (Cohen’s d = 0.90), surviving both Bonferroni and Benjamini–Hochberg correction. A secondary-test protocol on fresh class pools, feasible only for N ∈ {10, 20, 40} on CIFAR-100, shows that only 47.4% of positive-primary anomalies replicate in that regime, indistinguishable from chance: in the low-N regime where portability could be tested, CLEM has not acquired a fully portable update rule. Portability at the headline N = 100 condition cannot be measured with CIFAR-100 and remains an open question. The scaling pattern across the grid points toward conditions—larger meta-training class pools, larger latent dimension, longer training—and toward datasets large enough to test portability of the headline result, in which a more capable instance of the architecture may succeed more reliably.

Large Language Models (LLMs) are increasingly explored for public-sector tasks in which outputs must be reliable, auditable, and useful within domain-specific workflows. This the- sis explores the potential of, and develops a prototype for, a multi-dimensional benchmark- ing pipeline for evaluating locally deployed LLMs across summarization quality, semantic similarity, faithfulness-oriented scoring, demographic bias, medical bias classification, parse reliability, latency, and domain-specific quality assessment. The prototype is applied to locally deployed open-weight models in the 2.7B–20B range. A proprietary GPT-5-nano baseline, provided by the Norwegian Institute of Public Health (FHI), is included where comparable outputs are available. The central case study is AMSTAR-2 systematic review appraisal: 25 systematic review articles on screen time and child/adolescent development are evaluated against FHI expert reference ratings across all 16 AMSTAR-2 items and the overall confidence rating. The results show that no single metric provides an adequate picture of model suitability. In biomedical summarization, lexical and semantic metrics can reward vague summaries that contain little substantive information. In the bias experiments, occupation-related generation reveals systematic gender associations, while medical bias classification remains difficult. A re-scored run with a corrected prompt and an explicit None option shows persistent default- label collapse, with models often over-predicting one class. In the corrected AMSTAR-2 experiment, no model exceeds an article-independent majority-class accuracy baseline, so item accuracy is interpreted through chance-corrected agreement: gemma3:4b, mistral:7b, and gpt-oss:20b reach item-level agreement with the FHI expert ratings comparable to the available GPT-5-nano baseline, whereas GPT-5-nano performs best on overall confidence ratings. As the GPT-5-nano outputs were produced under different prompting and on a smaller article set, this is a bounded comparison rather than a controlled ranking. A key methodological finding is that structured prompts can introduce validity artefacts. In the initial AMSTAR-2 run, several models copied a concrete JSON example from the prompt rather than assessing the article. Response-variance analysis detected this template-copying behaviour, and a corrected placeholder-based prompt changed the model rankings. The thesis therefore argues that LLM evaluation for public-sector use should be task-specific, multi-dimensional, and explicitly validated for prompt sensitivity, structured-output reliabil- ity, and response variance. The evaluated local models appear promising as decision-support tools, but not reliable enough for autonomous quality appraisal without expert review.

AI systems built on tiny language models are being employed more and more for tasks like information retrieval, summarization, and addressing document questions. Many of the systems in use today depend on cloud-based services, which might have problems with latency, cost, privacy, and dependence on outside infrastructure. This thesis investigates the design and assessment of a local Small Language Model-based document question-answering system to respond to queries regarding uploaded documents while maintaining document processing on the user's own computer. This project presents the design, implementation, and assessment of a local document question-answering prototype. The system analyzes uploaded documents locally, breaks them up into text segments, indexes them using TF-IDF, and searches them using cosine similarity. It is based on a Retrieval-Augmented Generation architecture. The most pertinent document excerpts are then sent via Ollama to a locally running Small Language Model, which generates answers depending on the recovered context. To improve transparency, the system also provides source references and supports PDF, DOCX, and TXT files. The prototype gradually included optical character recognition using Tesseract to extract text from scanned or image-based document content. Additionally, the incorporation of a local vision model called llama3.2-vision enabled direct picture uploads and visual examination of produced PDF pages. As a result, in addition to the extracted text, queries concerning visible figures and graphics in documents might be asked. The system was assessed using specific materials and controlled questions in both Norwegian and English. The results show that the text-based RAG pipeline was the most dependable part of the system. It was useful for general document interpretation, goal-setting, information summarization, and responding to source-based inquiries. The examination also showed how chunk size affects retrieval quality since smaller chunks often produced more focused retrieval outcomes in the studied materials. While OCR facilitated the retrieval of text embedded in images, it did not provide full visual comprehension. The vision model could identify visual objects and figures, but it was not as precise as text-based retrieval and sometimes included information that were not immediately apparent. The findings also show how local hardware affects actual system performance. The prototype functioned on a laptop, but response generation became significantly faster when tested on a more powerful computer with more RAM and specialized graphics technology. This shows that the utility of local SLM-based systems depends on the hardware available, even if they may improve privacy and reduce dependency on cloud APIs. The thesis finishes by demonstrating how local Small Language Models might be useful for responding to document queries when combined with retrieval, OCR, and limited vision assistance. The system is best suited for text-based document interaction, while OCR and vision-based features should be viewed as useful but less dependable additions. Although there are trade-offs in terms of hardware performance, retrieval quality, setup complexity, and visual interpretation, the experiment shows that local AI systems may provide privacy-preserving document interaction.

Accurate classification of cystoscopic bladder cancer depends on subtle vascular patterns that small networks struggle to recover from RGB images alone, and reported accuracies in the field are often inflated by the absence of patient-level data splitting. A fast, generalizable classification model could shorten diagnosis and reduce cost. This thesis investigates whether combining datasets and supplying explicit vascular structure as an auxiliary input to a dual- stream classifier improves the classification of endoscopic bladder tissue under strict patient- level evaluation. The proposed pipeline combines an RGB backbone with a lightweight auxiliary stream pro- cessing a binary vessel mask and a Sobel edge map, fused through a channel-attention mod- ule. Three backbones (ResNet-50, ConvNeXt-Tiny, Swin-Tiny) are evaluated under an iden- tical warm-started, Optuna-tuned, five-seed protocol on the Lazo four-class tissue-grading dataset and the CystoDS dataset as binary ROI. Several findings emerge. On the four-class task, performance is limited primarily by data scarcity. The minority classes are barely learned on the Lazo dataset alone (macro F1 0.26– 0.53 across backbones), and merging Lazo with CystoDS into a single 174-patient dataset raises macro F1 to 0.60–0.68. It also recovers the previously collapsed classes, a gain a recipe- matched baseline attributes to the larger patient pool rather than to the training recipe. The auxiliary mask stream, by contrast, yields no consistent improvement on the four-class task. Its effect is small, changes sign across backbones, and is not a reliable difference, though it reliably reduces seed-to-seed variance. With the proposed pipeline, a 28M-parameter ConvNeXt-Tiny attains 0.924 macro-F1 on the CystoDS binary task, exceeding the published 197M-parameter Swin-Large baseline of Lee et al. by +0.068 at roughly one-seventh the parameter count.

The continuous growth in connected devices, Internet of Things (IoT) systems, and bandwidth-intensive applications has significantly increased the demand for wireless network capacity and performance. While modern cellular technologies such as Fourth Generation (4G) and Fifth Generation (5G) mobile systems provide improved spectral efficiency, low latency, and support for large numbers of users, expanding cellular infrastructure remains costly due to the limited availability of licensed spectrum. At the same time, WiFi technologies operating in the unlicensed spectrum have evolved to provide high-capacity wireless connectivity through dense and cost-efficient deployments. This has ignited interest in Carrier WiFi solutions, through which WiFi networks are integrated with mobile operator infrastructure to offload traffic from cellular networks. However, telecommunication infrastructures and Carrier WiFi solutions have traditionally been dominated by proprietary and closed ecosystems, limiting flexibility, interoperability, and accessibility for experimentation and innovation. This thesis showcases the design, implementation, and evaluation of an open-source Carrier WiFi architecture enabling Subscriber Identity Module (SIM)-based authentication for WiFi access. Our solution is primarily built based on Magma and integrates WiFi infrastructure with existing mobile Core Network (CN) components, supporting seamless interoperability between WiFi and cellular networks. By leveraging authentication mechanisms such as Extensible Authentication Protocol (EAP)-AKA and Passpoint, the system enables seamless onboarding and authentication of Commercial Off-The-Shelf (COTS) User Equipment (UE) using SIM credentials only, without requiring manual user configuration. The work further investigates differentiated Quality of Service (QoS) policies and subscriber-specific traffic handling within the Carrier WiFi network. Extensive experimental evaluation was conducted using multiple commercial UEs to assess authentication functionality, global end-to-end connectivity, policy enforcement, network performance, and inter-Radio Access Technology (RAT) handoff between mobile and WiFi networks. Through real-life experiments, we demonstrate the successful implementation of a open-source Carrier WiFi architecture supporting SIM-based authentication, differentiated service handling, and seamless access across WiFi and cellular networks while maintaining a satisfactory user experience.

In this thesis, we developed a Finite Element Method (FEM)-anchored Physics-Informed Neural Network (PINN) framework for transient water-quenching simulation of A356 aluminium. We based the study on the industrial quenching scenario of Mortensen et al. (2026). Our central question was: how many FEM solver calls can be replaced by PINN predictions while keeping the temperature error within engineering limits? We adopted Multi-Step Window Prediction (MSWP) as the core strategy. The FEM solver provides anchor temperature fields at selected time steps. The PINN predicts the field between consecutive anchors. The skip factor k controls the anchor spacing. We compared three Neural Architecture Search (NAS)-selected architectures: Bayesian/TPE (5×151, ReLU), NSGA-II (3×153, tanh), and NSGA-III (3×75, tanh). We conducted experiments across ten unique benchmark geometries in two studies. The canonical study covers eight domains: three 2D, four 3D, and the Thermal Fin, originally a 2D steady-state RBniCS benchmark that we extended to a 3D transient quenching scenario. A complementary adaptive-k study added Square and Flower as two new 2D geometries. On the adaptive-k study, we found 83–85% FEM saving with mean absolute error (MAE) below 2.5°C across all architectures. On the canonical 2D domains with fixed k, Bayesian/TPE stayed below 5°C through k=3 on all three domains. On the 3D domains, Bayesian/TPE remained acceptable at k=5 with MAE below 14°C. On the Thermal Fin, Bayesian/TPE with Fourier embedding reached k=5 at 12.55°C (ten-seed mean). NSGA-II and NSGA-III held through k=4. We demonstrated that FEM-anchored PINNs can reproduce transient quenching fields with 65–85% fewer solver calls within practical engineering accuracy.

AI systems built on tiny language models are being employed more and more for tasks like information retrieval, summarization, and addressing document questions. Many of the systems in use today depend on cloud-based services, which might have problems with latency, cost, privacy, and dependence on outside infrastructure. This thesis investigates the design and assessment of a local Small Language Model-based document question-answering system to respond to queries regarding uploaded documents while maintaining document processing on the user's own computer. This project presents the design, implementation, and assessment of a local document question-answering prototype. The system analyzes uploaded documents locally, breaks them up into text segments, indexes them using TF-IDF, and searches them using cosine similarity. It is based on a Retrieval-Augmented Generation architecture. The most pertinent document excerpts are then sent via Ollama to a locally running Small Language Model, which generates answers depending on the recovered context. To improve transparency, the system also provides source references and supports PDF, DOCX, and TXT files. The prototype gradually included optical character recognition using Tesseract to extract text from scanned or image-based document content. Additionally, the incorporation of a local vision model called llama3.2-vision enabled direct picture uploads and visual examination of produced PDF pages. As a result, in addition to the extracted text, queries concerning visible figures and graphics in documents might be asked. The system was assessed using specific materials and controlled questions in both Norwegian and English. The results show that the text-based RAG pipeline was the most dependable part of the system. It was useful for general document interpretation, goal-setting, information summarization, and responding to source-based inquiries. The examination also showed how chunk size affects retrieval quality since smaller chunks often produced more focused retrieval outcomes in the studied materials. While OCR facilitated the retrieval of text embedded in images, it did not provide full visual comprehension. The vision model could identify visual objects and figures, but it was not as precise as text-based retrieval and sometimes included information that were not immediately apparent. The findings also show how local hardware affects actual system performance. The prototype functioned on a laptop, but response generation became significantly faster when tested on a more powerful computer with more RAM and specialized graphics technology. This shows that the utility of local SLM-based systems depends on the hardware available, even if they may improve privacy and reduce dependency on cloud APIs. The thesis finishes by demonstrating how local Small Language Models might be useful for responding to document queries when combined with retrieval, OCR, and limited vision assistance. The system is best suited for text-based document interaction, while OCR and vision-based features should be viewed as useful but less dependable additions. Although there are trade-offs in terms of hardware performance, retrieval quality, setup complexity, and visual interpretation, the experiment shows that local AI systems may provide privacy-preserving document interaction.

Creating high-quality datasets is often tedious, expensive, and prone to errors. This thesis introduces a pipeline for creating a high-performance image classifier from noisy and imbal- anced data. The pipeline consists of a training dataset preparation and refinement stage integrating Segment Anything Model 3 (SAM 3) and Cleanlab. SAM 3 intelligently crops the images, and Cleanlab handles label errors and other forms of noise. Class imbalance is addressed during training by using weighted random sampling. The performance of the pipeline is demonstrated by creating a passerine bird species image classifier suitable for mobile phones and other resource-constrained devices. Noisy citizen science images from the Norwegian Species Observation Service, not intended for use in ML, are refined and used to train a MobileNetV4 classifier that achieves an accuracy of 89.82% on images depicting 94 species of passerine birds.

Automatic Speech Recognition (ASR) requires models that can process long acoustic sequences while preserving information across multiple temporal scales. Transformer-based systems achieve strong performance but rely on attention mechanisms whose computational and memory cost grows quadratically with sequence length. State Space Models (SSMs) provide an alternative sequence- modeling framework with recurrent structure and linear-time processing, making them a promising candidate for efficient speech transcription. This thesis investigates SSM-based ASR by designing, implementing, and evaluating a complete attention-free speech transcription pipeline built around SSM components. The pipeline uses SSM architectures at two levels: as acoustic encoders trained with Connectionist Temporal Classifi- cation (CTC), and as character-level language models (LMs) for shallow-fusion rescoring during decoding. Three encoder architectures are compared, a Simplified Selective State Space Model (S- SSSM), Mamba-1, and Mamba-3. The S-SSSM uses a diagonal gated linear recurrence with scalar per-channel state and bounded activations rather than the continuous-time discretization used in structured SSMs. The encoders are trained on LibriSpeech at 100 h, 460 h, and 960 h scales to evaluate both controlled architectural changes and full-scale system performance. Character-level n-gram models with Stupid Backoff smoothing are used as fixed-context controls for evaluating the contribution of the SSM-based LMs. The final system combines an SSM encoder with prefix beam search and a Mamba-based character-level LM. Building such a pipeline is not straightforward. The system replaces attention with recurrent SSMs in both the encoder and the LM, which means that the usual Transformer-based ASR setup cannot be followed directly. SSM models are also still developing, and Mamba-3 was released near the end of this research, so there was limited established guidance on how to use these models in ASR. In addition, there were no clear baselines for using character-level SSM-based LMs with shallow fusion. Therefore, this thesis compares the SSM-based LM with statistical n-gram models to better understand what the recurrent state contributes. The results show that S-SSSM achieves the lowest Word Error Rate (WER) among the tested encoders under the fixed-data experimental protocol, while Mamba-based encoders provide a sub- stantially more computationally efficient implementation path. Non-uniform initialization improves S-SSSM performance, but sensitivity experiments show that the benefit comes from initialization diversity rather than from a specific biologically motivated ordering. LM rescoring improves the final pipeline, and the Mamba LM remains beneficial in settings where the fixed-context n-gram control becomes less effective. The thesis contributes a compact and reproducible pure-SSM ASR pipeline, a controlled comparison of SSM encoder architectures, and an empirical analysis of initialization diversity and SSM-based LM integration for speech transcription. The final pipeline obtains 5.25% WER on LibriSpeech test-clean and 14.91% on test-other, corresponding to an unweighted mean WER of 10.08% across the two LibriSpeech test sets. Based on the Open ASR Leaderboard entries available at the time of writing, this result would place the system near rank 76 of 88 LibriSpeech entries. While not state of the art in absolute WER, the system is positioned as a candidate for the first pure-SSM Open ASR Leaderboard submission, with SSM components used for both acoustic encoding and language-model rescoring. Since the experiments in this thesis are limited to LibriSpeech, the process of full benchmark evaluation and submission is left for future work (Section 8.1).

Open-source implementation represents a promising tendency for private 5G network development and deployment. Representatively, several open-source 5GC projects have been developed in recent years for research and development purposes, making it possible to study, test, and evaluate real 5G network functions in flexible experimental environments. Based on free5GC, this thesis project designs, implements, and evaluates an end-to-end 5G system which supports both 3GPP access and untrusted non-3GPP access. The designed system contains one common network architecture and is implemented in two operation modes. Operation Mode-V is implemented as a virtual testbed using free5GC, UERANSIM, and N3IWUE, while Operation Mode-H is implemented as a hardware-based testbed using free5GC, OAI RAN, OpenWiFi, SDR hardware, COTS UEs, and N3IWUE. Through one working system which supports both operation modes, we showcase the operability and interoperability of both 5G and WiFi networks through one common core network platform. Furthermore, extensive experiments have been performed to evaluate connectivity, throughput, latency, packet loss, and QoE. The results show that both 3GPP access and non-3GPP access were successfully established through the common 5GC. In the final system, the VM hosting the free5GC CN was allocated 8 CPU cores and 8192 MB RAM, the highest measured DL throughput at the free5GC VM reached approximately 127 Mbps (Ookla), while the UE-side iperf3 measurement showed an average of approximately 117 Mbps for a single connected 5G UE with 0\% packet loss. In addition, the QoE evaluation demonstrated successful 2160p (4K) video streaming without buffering, confirming the capability of the implemented system to support bandwidth demanding user applications.

This thesis compares three approaches to document-level binary sentiment classification on the Norwegian Review Corpus (NoReC). The first is classical TF-IDF baselines with Logistic Regression and Linear Support Vector Classification. The second is a fine-tuned monolingual transformer encoder, NB-BERT-base. The third is zero-shot and four-shot prompting of three Llama 3 Instruct models at 1B, 3B, and 8B parameters. All three use the same train/dev/test split. Accuracy and macro-F1 are the main metrics, and the original 1–6 ratings are mapped to binary labels in the standard way: 1–3 as negative and 4–6 as positive. Fine-tuned NB-BERT-base and the LLM four-shot configurations are run with five seeds and reported as mean ± standard deviation. Fine-tuned NB-BERT-base reaches 0.8529 ± 0.0038 macro-F1, about 3.4 macro-F1 points above the strongest classical baseline. The best result overall comes from Llama-3.1-8B-Instruct with four in-context demonstrations: 0.9131 ± 0.0044 accuracy and 0.8777 ± 0.0036 macro-F1. That is about 0.025 macro-F1 above the fine-tuned encoder, a gap much larger than the observed seed-to-seed standard deviations on either side, and it is reached without task-specific training. Within the Llama 3 family, the 1B model is unstable across demonstration choices, with macro-F1 ± 0.043, and the 3B model is already strong in zero-shot. The 8B model is the only one that gets a clear gain from in-context demonstrations, and it stays almost unaffected by which four examples are used, with macro-F1 ± 0.0036. A chunk-and-pool re-evaluation of NB-BERT-base across the same five seeds, with [CLS] and [SEP] rebuilt per window, recovers part of the long-review penalty: mean macro-F1 on reviews longer than 1,000 words rises from 0.608 to 0.637, while the prompted LLM remains ahead at 0.702 on the same bin. A prompt-sensitivity check across three system-prompt variants leaves the main result unchanged. Per-example error analysis shows that the best fine-tuned and best prompted models disagree on about one in nine reviews and make different mistakes on different parts of the corpus. Consistent with this, a simple unweighted three-way majority vote across the strongest classical, fine-tuned, and prompted configurations beats any single member. The accompanying code repository contains the code, configurations, and pinned dependencies needed to reproduce the experiments.

The applicability of machine learning (ML) models for deployment on resource-constrained edge devices has become increasingly relevant with the growth of the Internet of Things (IoT) and embedded systems. Anomaly detection is particularly important in such environments, where reliable and efficient models are required to identify abnormal behavior in real time. However, strict limitations in memory size, computational capacity, and energy availability make the deployment of ML models on microcontroller (MCU)-based platforms challenging. This thesis investigates lightweight one-class classification (OCC) models for edge-based anomaly detection, with a focus on comparing three Tsetlin Machine (TM)-based approaches and a neural network (NN)-based autoencoder approach. TMs represent a promising logic-based learning paradigm, offering interpretable rule-based representations and potentially efficient inference for resource-constrained systems. Both model types are implemented and deployed on an ESP32-C6 MCU to evaluate their behavior under realistic hardware constraints. The models are evaluated on multiple datasets, including industrial supervisory control and data acquisition (SCADA) data, electrocardiogram (ECG) signals, and occupancy data, where the evaluation considers predictive performance, memory usage, inference time, and energy consumption. The results show that model suitability is highly dependent on the application context and the constraints of the deployment environment. While TM-based models offer advantages related to interpretability and compact logical representations, they currently lack the optimization maturity of TensorFlow-based NN deployments. As a result, the TensorFlow Lite for Microcontrollers (TFLM)-based models achieve stronger deployment efficiency in several experiments, while the TM-based approaches demonstrate potential for further improvement through dedicated embedded optimization. Overall, the results reveal a multidimensional trade-off between predictive performance and resource efficiency. This emphasizes the importance of context-aware model selection when designing embedded anomaly detection systems.