On 30th March 2022 all Russian participation in ICES was temporally suspended. Although the announcement of the suspension stressed the role of ICES as a “multilateral science organization”, this suspension applied not only to research activities but also to the ICES work providing fisheries advice for the sustainable management of fish stocks and ecosystems. On 9th December 2024 Russia announced its intention to leave ICES, and this came into effect on 9th December 2025. As a result, since 2022 the ICES AFWG has provided advice only for saithe, coastal cod north, coastal cod south, and golden redfish. Northeast Arctic (NEA) cod, haddock, beaked redfish, Greenland halibut, and capelin assessments have been conducted outside of ICES in a bilateral Joint Russian-Norwegian Working Group on Arctic Fisheries (JRN-AFWG) and should not be considered as ICES advice. Although this work has been conducted independently of ICES, the methodologies agreed at ICES benchmarks and agreed HCRs (Harvest Control Rules) have been followed in providing this advice. The chapter numbering in this report is not continuous (Chapter 2,5 and 7 are missing). This is done in order to be consistent with the chapter numbering in ICES AFWG reports. In 2026 we are giving annual advice for NEA cod, NEA haddock, beaked redfish, and a two-year advice for Greenland halibut. The beaked redfish typically has a two-year advice, however this stock is in the process of a benchmark model revision. Therefore, we issue single year advice this year, with the intention of reverting to two-year advice next year following the results of the benchmark. In 2025 TAC quota was set above advice for cod, haddock and Greenland halibut. In 2026 the quota was above advice for cod and Greenland halibut. We note that cod, in particular, is in an extremely sensitive situation, with projected future recruitment to the fishery estimated to be very low in the medium term. Therefore, any future TAC over advice poses a real risk of reversing the projected modest upturn in the stock. Advice on fishing opportunities for NEA cod The NEA cod stock is continuing to decline following a prolonged period of moderate to poor recruitment, and is currently below Bpa but above Blim. The stock is forecast to rise next year as moderate recruitment enters the fishable stock. However, projections for recruitment in the following years are very low. As a result, the stock is projected to stabilize a little above Bpa , provided the advice is followed. The yearclasses entering the fishery now are therefore critical to the medium-term development of the stock, and overfishing these would pose a risk of an extended period of stock below precautionary levels. There was a model revision this year, to better account for the different levels of uncertainty at different magnitudes of catches. This is the same method as used for haddock and now the ICES NEA saithe assessment and has improved both the fit to the data and slightly improved the model retro. This revision has slightly revised up the stock estimate. Following the agreed HCR, the advice for 2027 is that catches should be no more than 312 667 tonnes, up from an advice of 269 440 tonnes and a quota of 285 000 tonnes in 2026. Provided that this advice is followed, then projections indicate that the stock should start to rise slightly and then stabilize at moderately low but safe levels. It should be stressed that this quota advice is higher than would be expected from the current state of the stock (which would give a quota advice of 272 528 tonnes) because the HCR accounts for the projected increase in the stock. F urther quota above advice could be expected to delay any recovery for this stock, and potentially lead to further decline. Advice on fishing opportunities for NEA haddock Following a period of low recruitment, the stock is rising as a result of relatively good yearclasses in 2021 and 2022 currently entering the fishery. Following the agreed HCR, the advice is that catches in 2027 should not exceed 180 336 tonnes, up from an advice and quota of 153 293 tonnes in 2026. Provided that these incoming yearclasses are not heavily caught at small sizes, then then this should lead to a continued increase in stock and catches as long as the advice is followed. In recent years there has been a rise in the catch of small haddock, and if this is not curtailed then there is a risk that a large part of the incoming yearclasses could be fished before reaching a size to give optimum yield. Advice on fishing opportunities for Greenland halibut The stock is on a downward trajectory towards Bpa. However, improved recruitment is now entering the adult and the fishable biomass. This gives a potential to reverse the downward trend provided fishing pressure is kept at appropriate levels. The total biomass is project to rise slightly in 2026 if the advice is followed, although it will take several more years for the female SSB to begin to rise. There has been an upwards retrospective revision for the Greenland halibut stock, as a result of which the stock is now assessed as being above Bpa, compared with the 2024 assessment which projected the stock to fall below Bpa in the course of 2024. The advice is that catches in 2027 should be no more than 19 610 tonnes, and catches in 2028 should be no more than 19 914. These are higher than the advice of 12 431 tonnes for 2025 and 14 891 tonnes for 2026, but lower than the TAC (19 000 for both years) and lower again than the estimated catches of 20 652 tonnes in 2025. We stress that we present a total quota advice, which includes catches in international waters and a small area of UK waters. This stock has a history of quota and catches being set above advice, and then catches above the TAC, which has led to the decline of the stock. Advice on fishing opportunities for beaked redfish The stock is at a high level, with SSB relatively stable. Recent recruitment is slightly down on the peak levels around 2020, but is still at good levels and well above the low levels seen in the late 1990s and early 2000s. The stock is fished at a moderate fishing pressure in line with advice. The current model is considered rather uncertain (largely due to lack of a coherent survey coverage and limited age data), with a high degree of annual retrospective adjustments, and there is no accepted precautionary Fmsy level. A benchmark revision is in process. Therefore, a single year’s advice is presented this year, instead of the typical two-year advice. The aim is to revert to two-year advice next year. The catch advice is no more than 67 090 tonnes in 2027, compared to an advice of 67 191 tonnes in 2025 and 69 177 tonnes in 2026. TAC has historically been set following advice. It should be noted that because there is not an agreed precautionary Fmsy value for this stock, a “status quo F” approach is used to set the advice. It is therefore possible that the significant changes in the quota advice could occur following the benchmark once an appropriate target fishing mortality can be evaluated.

We introduce a new rule-based algorithm called GEOPARD, which models shoreface deposits by following geological principles. In this algorithm, the outcomes of geological processes are represented as rules that are integrated into the core of a standard geostatistical modeling framework. The GEOPARD builds on the stochastic object-based facies modeling technique and incorporates a Bayesian framework for conditioning to data and reducing uncertainty. This paper outlines the geological prior model of the GEOPARD algorithm, which generates facies geometries and controls object placement using geological rules. Specifically, the algorithm builds up a parasequence by stacking a succession of prograding shoreface bedsets, bounded by small-scale hiatus, until a final point of maximum shoreline advance. Model parametrization closely follows the geological conceptual model. The rules are implemented as a series of fully automated modeling steps, mapping the wide variety of facies geometries typically associated with shallow-marine deposits. The functionality of GEOPARD is demonstrated through a series of scenarios, including the reproduction of features observed in an outcrop analogue and benchmarking against the truncated Gaussian simulation method. Key modeled features include sand-body thickness, lateral extent of facies, overall parasequence geometry, and the spacing and dip of bedset bounding surfaces.

PCube+ represents posterior lithology-fluid class probabilities as a discrete-time Markov chain at the sampling resolution of the input seismic data. This note presents a method for constructing a finer posterior chain without upsampling the seismic input or rerunning the inversion at higher resolution. Posterior marginals are interpolated in logit space using Catmull-Rom splines. Fine-step transition matrices are approximated from the coarse transitions while preserving legal transitions, and are subsequently corrected to make the interpolated marginals reachable. Derived classification, thickness, and elastic-property outputs are then calculated from the upsampled posterior. Examples demonstrate smoother classification and elastic-property maps, together with reduced sample-interval bias in a synthetic thickness estimate. Synthetic seismic and residual RMS values remain comparable to those obtained from the original results, indicating that the procedure does not significantly alter the seismic response. The method introduces no new seismic information and should not be interpreted as equivalent to performing PCube+ inversion at a finer resolution. It provides a computationally inexpensive means of improving visualization and the sampling of posterior-derived quantities.

Data spaces and data marketplaces are characterized by dynamic ever-changing availability of decentralized data sources, users, and accesses where data can be access, can be granted and revoked at any time, Data sources may become faulty or unavailable over time, and new data sources might be added. This dynamicity requires the creation of an Authentication and Authorization Infrastructure (AAI) designed for resilient and sustainable applications. It must offer great flexibility and adaptivity providing secure and dynamic authentication, authorization, and access control services. While user management policies provide the foundation for identity and access management (IAM) systems, a dynamic policy management process enables policy adaptability for dynamic IAM scenarios. This process utilizes key performance indicators to detect and refine policies, facilitating the sustainable adoption of Self-Sovereign Identity (SSI). Therefore, we have developed a dynamic and adaptive authentication and authorization infrastructure (DA3I), considering the requirements of service providers to manage dynamic target platforms, decentralized and federated data sources, users, and access points. DA3I enables robust authentication and authorization within a multi-domain resources environment offering a resilient, sustainable, secure and dynamic access control system. We have also highlighted the significance of this study for offering long-term stability and protection to the Belt and Road Initiative (BRI) initiative and elaborated the characteristics and challenges of applying dynamic and adaptive trust in federated systems. Moreover, we presented a broad range of applications of the proposed DA3I approach in various critical systems such as healthcare, transportation, energy and industrial control systems.

The EU-CIP project has developed a methodology for continuous data collection and desk research, aimed at consistently gathering and analyzing data from critical infrastructures. This methodology facilitates the ongoing collection, analysis, curation, extraction, and presentation of information and insights about critical infrastructure protection and resilience (CIP/CIR). In this book chapter, we present this methodology and analyze the outcomes of two questionnaire surveys conducted within the EU-CIP consortium (i) internal stakeholders feedback survey: evaluating technological gaps, needs, and innovations in CIP/CIR, and (ii) survey on applying the developed methodology for capability and technology analysis in CIP/CIR. The survey results reveal some critical capability needs identified as lacking including enhanced adaptability, reduced response times, improved detection and impact assessment, and tools for managing cascading effects between entities and across borders. The results also revealed a preliminary list of capability gaps that align with the identified needs. These include inadequate automation, insufficient control over interconnectedness, poor alignment of resilience indicators, the absence of standardized stress-testing procedures, classification issues with IoT devices, scalability challenges in mitigating DDoS attacks, and gaps in emergency management processes. Addressing the identified capabilities needs and gaps is necessary for CIP/CIR. These insights will contribute to refining the strategies for the EU-CIP initiative, with a focus on practical and innovative solutions for CIP/CIR.

This open access book provides an exploration of how technology and innovation are transforming the field of Critical Infrastructure Resilience.

The problem of assuring cybersecurity and resilience of IT infrastructures and services has been challenging both the research community and practitioners for decades. Despite the joint effort by the industry, academia, and government, the society has been constantly suffering from cyber breaches and incidents. This prompts the following research question: What would be a new kind of cybersecurity architecture that can potentially enhance the cybersecurity and resilience of IT infrastructures and services? This paper explores a novel architecture as well as the associated mechanisms that can be incorporated into the architecture to lead to resilient IT infrastructures and services. The innovation of the proposed architecture includes: (i) it is based on the core concept of Digital Twins, which, to our knowledge, has not been systematically explored in the context of IT cybersecurity; (ii) it is mapped to the NIST Cybersecurity Framework (CSF) 2.0 functions, which is deemed important because practitioners often use CSF 2.0 to guide their cyber defense practice; (iii) the architecture offers a seven-layer design, which builds on top of the existing IT cybersecurity mechanisms, such as intrusion detection, firewalls, and cryptographic solutions, and emerging techniques such as zero-trust.

The healthcare sector profoundly influences humanity, society, and the global economy, as every individual requires healthcare services at some point in their life. This demand is particularly significant with the growing elderly population, driven by continuous advancements in medical science and technology. However, cyberattacks targeting healthcare infrastructures and services are increasing, as attackers can derive considerable financial gains from successful breaches. This trend underscores the need for research focused on achieving resilient healthcare infrastructures and services. In this context, an innovative architecture, introduced at SUNRISE 2024 and termed the Adaptive Cognitive Digital Twins (ACDT) architecture, is proposed to enhance the resilience of healthcare systems. This paper investigates methods for modeling the ACDT architecture, explores mechanisms that can be integrated to implement resilient healthcare infrastructures and services, and identifies suitable metrics for evaluating their resilience. Consequently, this work constitutes an important step toward translating the conceptual ACDT architecture into real-world applications.

Presentasjon av hvordan SCERTS-rammeverket ble tilpasset og anvendt som forskningsverktøy i ROSA-prosjektet. Innlegget beskriver ulike roller den sosiale roboten Robbie kan ha i læringsmiljøet, blant annet som samhandlingspartner, modelleringsverktøy, læringsverktøy og samtaletema, samt hvordan SCERTS ble brukt til å analysere samhandling og læringsstøtte

Presentasjon av ROSA-repertoaret, inkludert læringsaktiviteter, lærergrensesnitt, elevgrensesnitt, innhold fra Cyberbook og funksjonalitet knyttet til den sosiale roboten NAO. Innlegget beskriver hvordan teknologiske komponenter og pedagogisk innhold ble kombinert til et fleksibelt repertoar for bruk i spesialundervisning, samt erfaringer fra implementeringen i skolehverdagen

Presentasjon av hovedresultater fra ROSA-prosjektet. Innlegget omfatter utviklingen av ROSA-løsningen, tilpasning og bruk av SCERTS som forskningsverktøy, erfaringer fra pilot- og hovedstudier, samt analyser av kvalitet og utvikling i elev–robot-interaksjoner. Resultatene viser både positive læringseffekter og betydningen av tekniske og organisatoriske rammebetingelser for bruk av sosiale roboter i spesialundervisning.

Videoopptak fra sluttseminaret for ROSA-prosjektet (Robotstøttet læring for barn med autisme), gjennomført ved Frydenhaug skole 13. april 2026. Seminaret presenterer erfaringer, forskningsresultater og metodiske refleksjoner fra prosjektet, med bidrag fra forskere, lærere og samarbeidspartnere. ROSA var et forskningsprosjekt ledet av Norsk Regnesentral i samarbeid med Universitetet i Sørøst-Norge, University of Birmingham, Frydenhaug skole, Drammen kommune, Cyberbook AS og Innocom AS.

Objectives To examine screening mammograms assigned high-risk scores by two artificial intelligence (AI) models, 2 and 4 years prior to screen-detected cancers. Materials and methods This retrospective study was based on data from 130,031 screening examinations performed in BreastScreen Norway (2008–2018), processed by two AI models: a commercial model (A) and an in-house model (B). The study sample included women with screen-detected cancer, and two prior consecutive screening examinations assigned the highest 10% of AI risk scores. Two radiologists conducted an informed review of three consecutive screening mammograms per woman, 4 and 2 years prior to, and at diagnosis. Descriptive statistics assessed the location of AI markings, mammographic features, histopathological tumor characteristics, and visibility of malignancy preceding diagnosis. Results Model A and B assigned high-risk scores at both prior mammograms for 43 and 47 cases, respectively, with 29 cases selected by both, yielding 61 sets of three consecutive examinations. AI markings corresponded to the cancer location in at least one view in 61% and 57% of cases 4 years prior to diagnosis for models A and B, respectively, while radiologists classified 89% as true negative or minimal sign non-specific. At diagnosis, spiculated mass (28%) and density with calcifications (20%) were the most frequent mammographic features but were absent 2 and 4 years prior. Conclusion The AI models identified breast cancer in a substantial part of screening mammograms 2 and 4 years preceding diagnosis, but only a few demonstrated suspicious findings according to radiologists. Key Points Question Understanding the correspondence between AI markings, cancer location, and mammographic features is important to evaluate AI models’ potential to enhance early breast cancer detection. Findings AI markings corresponded to the cancer location 4 and 2 years prior to diagnosis, with evolving mammographic features, yet radiologists classified most examinations as negative. Clinical relevance The discrepancy between AI-identified breast cancers years before diagnosis and radiologists classifying the same cases as negative highlights both AI’s potential and the challenges of how best to implement it to enhance early detection.

People have different views on the current level of intelligence in AI systems, the likely speed of further progress, and the societal impact AI may have in the future. This has contributed to a polarized debate, with narratives ranging from “AI will give us some useful tools”, via “Many jobs will disappear” and “We will all be rich”, to “AI will take over and kill us all”. In this presentation, I will map the main narratives about AI risk and benefit using a railway metaphor in which human extinction is the final destination. The different stops along the way represent arguments for more favourable outcomes. The goal is not to argue for the likelihood of any particular scenario, but to structure the discussion of these narratives in a systematic way.

Hyperbolic representation learning has shown compelling advantages over conventional Eu- clidean representation learning in modelling hierarchical relationships in data. In this work, we evaluate its potential to capture biological relations between cell types in highly multiplexed imaging data, where capturing subtle, hierarchical relationships between cell types is crucial to understand tissue composi- tion and functionality. Using a recent and thoroughly validated 42-marker Imaging Mass Cytometry (IMC) dataset of breast cancer tissue, we embed cells into both Euclidean and Lorentzian latent spaces via a fully hyperbolic variational autoencoder. We then introduce an information-theoretic framework based on k-nearest neighbour estimators to rigorously quantify the clustering performance in each geom- etry using mutual information and conditional mutual information. Our results reveal that hyperbolic embeddings retain significantly more biologically relevant information than their Euclidean counter- parts. We further provide open-source tools to extend Kraskov-Stögbauer-Grassberger based mutual information estimation to Lorentzian geodesic spaces, and to enable UMAP visualizations with hyper- bolic distance metrics. This work contributes a principled evaluation method for geometry-aware learning and supports the growing evidence of hyperbolic geometry’s benefits in spatial biology. Code is available at: https://github.com/youssefwally/FlatlandandBeyond

Cognitive Digital Twins (CDTs) are virtual models of physical systems that integrate cognitive functions with real-time Internet of Things (IoT) data to support simulation, decision-making, and cyber resilience. In domains such as healthcare and smart cities, CDTs enable advanced capabilities but also introduce significant privacy and security risks, especially due to continuous, real-time data exchange and automated decision-making. This paper presents DC-TWIN, a dynamic consent-enabled CDT framework that addresses the limitations of static, one-time consent models in real-time, data-intensive environments. DC-TWIN introduces a multi-layered architecture consisting of: (i) a multi-layered architectural stack, including user control, policy management, dynamic trust, and data governance layers, that supports compliance monitoring, risk-aware user interfaces, consent history tracking, federated identity and role binding, and adaptive trust modeling, and (ii) a dynamic consent reasoning engine that uses contextual signals (e.g., user role, device status, network activity) and human factors (e.g., cognitive load, trust calibration, fatigue) to assess data access requests in real time, issuing granular decisions (grant, deny, prompt) or escalating for clarification. We highlight key use cases in healthcare, welfare technologies, and smart cities. The framework empowers users with real-time, contextual control over how their data is accessed, shared, and reused through adaptive interfaces and personalized consent mechanisms. By integrating dynamic consent reasoning, trust calibration, and continuous feedback loops, DC-TWIN supports transparent, compliant, and user-aligned data governance. It contributes to the secure and ethical deployment of CDTs by reinforcing user autonomy, enhancing risk communication, and enabling responsive consent management in critical domains such as healthcare.

The cyberattack surface in critical sectors is expanding due to the rapid proliferation of Internet of Things (IoT) devices. Artificial Intelligence (AI) models, such as Deep Neural Networks (DNNs) and Convolutional Neural Networks (CNNs), offer promising capabilities for detecting and classifying cyber threats. However, these models often struggle to generalize to previously unseen attacks after deployment. This study investigates how well different AI techniques can generalize to such novel threats in the presence of class imbalance. We evaluate three data balancing strategies: Generative Adversarial Networks (GAN), Synthetic Minority Over-sampling Technique (SMOTE), and class weighting. Experimental results indicate that DNNs outperform CNNs when provided with identical input data. While each balancing method has distinct advantages and trade-offs, the highest multiclass accuracy of 81.16 % was achieved by a DNN using GAN-augmented data for the previously seen attack types. The best performance on unseen attacks was achieved by a DNN trained with SMOTE, yielding a multiclass accuracy of 51 % among eight classes. The binary classification (benign vs. malicious) results were satisfactory, with DNN using GAN-augmented data achieving an accuracy of 99.20 %. These findings highlight the importance of not only separating data into training and test splits, but also incorporating a “seen vs. unseen” evaluation strategy.

In this work, we explore particle learning strategies for the joint estimation of static parameters and latent states within conditional sequential Monte Carlo (CSMC) algorithms. Building on this idea, we propose the p(parameter)-CSMC algorithm, which incorporates both parameter learning and ancestor sampling, leading to much better mixing properties compared to Gibbs sampling in settings where strong internal correlations may challenge effective exploration. We include an application to the estimation of weights in a branching process model against synthetic data and show that, in this setting, performance is dramatically enhanced, with substantially faster mixing and markedly reduced autocorrelation compared with standard particle Gibbs implementations.

Pre-trained encoders produce semantically rich latent embeddings, which, however, may expose unintended information through malicious inference or exploitation. We propose SEAL, a framework that mitigates embedding leakage by disrupting latent representations based on information-theoretic principles. It reduces the risk of potential misuse while enabling controlled reconstruction for trusted users. SEAL learns to encode controlled perturbations by minimizing the Matrix Norm-based Quadratic Mutual Information (MQMI) functional between original and perturbed embeddings within a hyperspherical latent space. Meanwhile, a private decoder, jointly trained with the SEAL encoder, is trained to reconstruct the original data that is accessible only to authorized users under an access-controlled setting. Extensive experiments on vision and text datasets demonstrate that SEAL reduces latent leakage, weakens the effectiveness of evaluated inference attacks, and preserves reconstruction under the considered setting.

Background Radiomic models from multiparametric MRI can characterize tumor infiltration within the non-enhancing peritumoral region but remain insufficiently compared with diffusion and perfusion. This study assessed concordance between voxelwise radiomic predictions and these physiological modalities and evaluated prognostic value of infiltration metrics in two external cohorts. Methods UPenn-GBM and UCSF-PDGM datasets were analyzed. Voxelwise radiomic classification generated peritumoral infiltration-probability maps from standard MRI. Fractional anisotropy (FA), dynamic susceptibility contrast–derived relative cerebral blood volume (DSC-rCBV; UPenn), and arterial spin labeling–derived relative cerebral blood flow (ASL-CBF; UCSF) were compared between peritumoral regions classified as high- versus low-infiltration. Radiomic metrics included infiltration burden (voxel fraction exceeding predefined probability thresholds) and radial extent (normalized maximum distance from enhancing margin sustaining high infiltration probability) were quantified, and survival assessed using univariable Cox models. Results Across 872 subjects, high-infiltration regions showed significantly lower FA (median difference: UPenn −0.232; UCSF −0.226; both p < 0.001) and higher perfusion (median difference: UPenn DSC-rCBV + 0.282; UCSF ASL-CBF + 0.565; both p < 0.001) compared with low-infiltration regions. Infiltration burden at the 0.50 threshold demonstrated prognostic value (UPenn hazard ratio (HR) 2.758, 95% confidence interval (CI) 1.189–6.396, p = 0.018; UCSF HR 21.277, 95% CI 6.024–71.429, p < 0.001). Radial extent was also associated with survival (UPenn HR 2.371, 95% CI 1.215–4.625, p = 0.011; UCSF HR 4.405, 95% CI 1.695–11.494, p = 0.002). Conclusions Voxelwise radiomic infiltration mapping from standard MRI aligns with diffusion and perfusion abnormalities and yields prognostic value. These metrics highlight the role of structural radiomics for characterizing non-enhancing infiltrative spread in glioblastoma.

The Norwegian Centre for Research-based Innovation, Visual Intelligence, advances deep learning in marine science. The work focuses on analyzing multifrequency echosounder data to support ecosystem and fisheries management. To address limited labeled data for identifying sand eels in the North Sea, a semi-supervised learning method was developed that combines labeled and unlabeled data, significantly improving accuracy (Choi et al., ICES JMS 2021). Building on this, the method was extended to semantic segmentation (Choi et al., IEEE J. Ocean. Eng. 2023), enabling detailed classification of acoustic signals while reducing the need for costly annotations. More recently, foundation models were explored to tackle challenges like changing conditions in marine environments. By aligning these models with echosounder data and using semantic tokenization, they achieved strong performance with minimal labeled data (Choi et al., NAIS 2025). These innovations highlight the transformative role of AI in exploring and understanding the underwater world.

Arctic and Subarctic fisheries face increasing pressure as traditional practices struggle to keep pace with shifting ocean conditions. These challenges call for intelligent, adaptive systems to support sustainable monitoring and management. This session explores how Artificial Intelligence (AI) can contribute to the future of fisheries science by bridging algorithmic innovation with practical application and fostering interdisciplinary exchange. We invite papers that demonstrate how AI, ranging from machine learning and computer vision to emerging LLM-based agentic systems, can support fisheries research and regulation. Relevant applications include acoustic data interpretation for species identification, improved stock assessments, vessel activity monitoring, and adaptive regulatory strategies. Submissions addressing challenges such as limited computational resources, sparse data availability, or operational constraints in remote polar environments are particularly encouraged. The session also highlights the importance of engaging Indigenous knowledge holders and coastal communities in AI development. We especially welcome co-designed frameworks that promote fairness, transparency, and local agency in contexts where environmental data and governance intersect. Through this session, we aim to bring together diverse perspectives that advance responsible and locally grounded uses of AI for sustainable marine stewardship.

Utviklingen i innovasjonsprosjektet «landsdekkende myrdatasett» har vært preget av både teknologiske fremskritt og utfordringer knyttet til å etablere en landsdekkende KI-modell for kartlegging av myr og våtmark. Presnetasjonen tar for seg utviklingen og testingen av ulike modellutgaver, og hvordan de kan kombinere dem for å lage et mer robust og presist resultat, samt innsikt i testresultatene som foreligger.

We present a web application for controlling NAO V5 and NAO V6 robots using Hypertext Transfer Protocol Secure (HT TPS). The application was designed for a study in a special school. The school staff have found the application easy-to-use and versatile, and it may be useful to other researchers or people interested in controlling NAO. The HT TPS constraint and the locked-down nature of NAO introduced additional development challenges, and the solutions to these challenges are worth sharing with the HRI community. We present characteristics of the web application, implementation details, how it has been set up, and how to use it. Although the application is usable in its current form, there are still things that can be improved to make the application more useful in other contexts. We therefore document how the remote control can be extended and potential starting points for improvement.

Human-robot interaction (HRI) programmers often struggle with operating older robot hardware due to the short support period provided by manufacturers and difficulties integrating modern software solutions. This paper introduces the ZTL Task Library (ZTL), a lightweight communication framework and protocol designed to decouple robot hardware from the operating platform via socket communication, thereby increasing robot lifetime. We present a task-based communication protocol facilitating the co-design of robot behaviours with non-programming experts. Our approach has been shown across different platforms to effectively mitigate incompatibilities between middlewares, simplifying control and usability, allowing for simultaneous addressing of multiple devices.

Industrial Control Systems (ICSs), which are essential components of critical infrastructures, are inherently complex and vulnerable to cyberattacks. Advanced Persistent Threats (APTs) that target these systems are multi-stage, coordinated attacks that can lead not only to information loss but also to physical damage and loss of life. Traditional threat modeling approaches fall short in adapting to the dynamic nature of ICSs, necessitating new methodologies to predict and prevent such complex attacks. This work presents a digital twin-assisted dynamic threat modeling framework for ICS environments. The framework leverages a knowledge graph that integrates system data and cyber threat intelligence to predict potential attacks. In addition, the digital twin environment enables the validation of mitigation strategies before deployment in the physical system, while also supporting adaptive response and real-time mitigation. To predict the attacker’s next move, we propose a Relational Graph Convolutional Network (RGCN)-based model that utilizes enriched relational data such as tactics, campaigns, groups, techniques, and assets. The proposed RGCN model achieves a recall of 0.887, an F1-score of 0.893, and an AUC of 0.957 in predicting potential attack sequences. These results demonstrate that the model provides reliable and well-balanced predictive performance.

Å kompensere mennesker som bidrar med originalt innhold, handler ikke bare om rettferdighet. Det er også en investering i et bærekraftig, digitalt økosystem.

ntroduction: The increasing incidence of endometrial cancer (EC) requires an extensive search for novel preventive tools and early intervention approaches. However, the development of reliable predictive models is impossible without knowledge of genetic alterations prior to diagnosis. In this work, we aimed to establish whether known EC risk factors are associated with peripheral blood gene expression changes in a prospective design and whether such associations differ between women who later developed EC and matched controls. Methods: First, we selected variables (parity status, lifetime number of years of menstruation, coffee consumption, body mass index (BMI), age at menopause, use of oral contraceptives) that were shown to have an impact on EC risk in a large prospective cohort (165,000 women). Next, using BeadChip microarray technology, we tested the association between these variables and gene expression profiles in RNA extracted from mixed circulating immune cells in a nested case-control study (79 case-control pairs) of women from the NOWAC postgenome cohort. Lastly, we undertook a gene set enrichment analysis (GSEA). Results: At overall gene expression level, we found no difference between the EC cases and controls. The introduction of parity status into the statistical model revealed changes in the expression of 1,379 genes in the controls, while we did not observe any expression changes in the cases. Twenty-seven genes were associated with BMI increase in the controls, whereas there was no association observed between changes in BMI and gene expression in women with EC. In GSEA, 2,407 significantly enriched gene sets were attributed to a parity increase among cancer-free women. Conclusion: In this study, we found that an increased number of parities has a life-long effect on the gene expression profile in the peripheral blood of women who never developed cancer. In contrast, in women who were diagnosed with EC later in life, neither multiparity nor elevated BMI showed a significant association with gene expression patterns. However, given the modest sample size and exploratory nature of the study, these findings should be verified in larger cohorts.

Abstract Objectives To investigate mammographic features associated with high artificial intelligence (AI) risk scores as provided by two AI models applied to screening mammograms. Materials and methods This retrospective study included 130,031 screening mammograms from 42,371 women attending BreastScreen Norway, 2008–2018. Two AI models (A and B) developed for cancer detection on screening mammograms were applied. An informed radiological review was conducted for mammograms within the highest 5% of AI risk scores by both models in two study samples: (1) High AI risk score, but no breast cancer detected within 6 years ( n  = 120), and (2) High AI risk score in mammograms with screen-detected cancers ( n  = 120). Mammographic density (BI-RADS a–d), features (mass, spiculated mass, asymmetry, architectural distortion, calcification alone, and density with calcification), and radiologists’ interpretation scores (1–5) were analyzed descriptively. Results Mammographic density was higher in sample 1 compared to sample 2 (BI-RADS d: 11% vs 3%, respectively). In sample 1, calcifications alone were the most frequent AI-marked feature (model A: 72%; model B: 68%), predominantly with amorphous morphology and a cluster distribution, and 76% were interpreted as benign by the radiologists (interpretation score 1). In sample 2, a spiculated mass was the most frequent mammographic feature among the screen-detected cancers (29%). Conclusion Mammograms assigned high AI risk scores exhibit distinct features depending on screening outcome. Systematic characterization of these features may help refine AI thresholds, improve specificity, reduce AI false-positive findings, and decrease the recall rate in breast cancer screening. Key Points Question Knowledge about mammographic features associated with high AI risk scores is essential for distinguishing cancer from non-cancer cases. Findings Calcifications were the dominant feature in non-cancers in screening mammograms with high AI risk score, whereas spiculated mass was the most frequent feature among cancers. Clinical relevance Calcifications in non-cancer screening mammograms with a high AI risk score were frequently interpreted as benign or probably benign by radiologists. This knowledge may help refine AI thresholds and thereby improve specificity and reduce false-positive results in mammographic screening. Graphical Abstract

Objective: Tremor symptoms in Parkinson’s disease (PD) are challenging to assess due to low resolution and subjectivity from standard clinical scales. To address this, wearable devices have been used, but algorithms have been relying on controlled or limited activity conditions. Our objective is to create a context-independent metric quantifying tremor in free-living conditions to bridge the gap between biomedical engineering and the PD field. Methods and Procedures: We designed an algorithm which computes a tremor index (TI) from accelerometer data, collected via the Empatica E4 worn on the wrist by home dwelling people with PD. For validation, we use a within-participant design, comparing the TIs of the most and least tremor-affected hand. We included seven participants with unilateral tremor, monitored for two weeks each. The algorithm is able to compute TIs for a set of frequencies identified in literature as associated with different tremor types (3–12 Hz), over adjustable sampling time windows. Results: We show that the most tremor-affected hand yields a higher TI than the other hand for frequency sets that are individual to each person, in particular around 5-6 Hz where rest tremor typically occurs. We find that we can disambiguate tremor across 3-12 Hz from general movement and resting states. The number of frequencies with inter-hand separation correlate with the MDS-UPDRS part III tremor items. Conclusion: The designed tremor quantification algorithm can quantify tremor symptoms over time for people with PD and can be used to identify the individualized frequency ranges where these movements happen, in free-living conditions.

Generativ kunstig intelligens er imponerende, men ikke alltid så nyttig til å løse industrielle problemer.

An increased focus on the blue economy across coastal African countries requires effective strategies for reducing marine resource conflicts to achieve goals of sustainable, equitable ocean development. We created a spatial database documenting marine resource conflicts (2008–2018) and conducted an expert survey to analyze patterns in conflict types and how they relate to actors, drivers, and resolution. Our findings indicate that 73% of conflicts were associated with access disputes and 28% were between non-fisheries sectors. National governments, small-scale or industrial fishers, and state enforcement agents were the most frequent actors. Illegal fishing, inequitable benefit distribution, and inadequate regulations were commonly reported conflict drivers. Less than one third of conflicts were resolved, but increased governance was cited as important for resolution. These results suggest policymakers may need to focus on access and benefit sharing issues and increase engagement of key actors in governance processes to realize blue economy ambitions.

Text sanitization is the task of redacting a document to mask all occurrences of (direct or indirect) personal identifiers, with the goal of concealing the identity of the individual(s) referred in it. In this paper, we consider a two-step approach to text sanitization and provide a detailed analysis of its empirical performance on two recently published datasets: the Text Anonymization Benchmark (Pilán et al., 2022) and a collection of Wikipedia biographies (Papadopoulou et al., 2022a). The text sanitization process starts with a privacy-oriented entity recognizer that seeks to determine the text spans expressing identifiable personal information. This privacy-oriented entity recognizer is trained by combining a standard named entity recognition model with a gazetteer populated by person-related terms extracted from Wikidata. The second step of the text sanitization process consists in assessing the privacy risk associated with each detected text span, either isolated or in combination with other text spans. We present five distinct indicators of the re-identification risk, respectively based on language model probabilities, text span classification, sequence labelling, perturbations, and web search. We provide a contrastive analysis of each privacy indicator and highlight their benefits and limitations, notably in relation to the available labeled data.

When a database system is decommissioned, retaining its data, structure, and query capabilities is often crucial for future restoration and access. The Immortal Database Access (iDA) solution focuses on preserving decommissioned databases along with their stored information and retrieval functionalities. Building upon the Immortal Virtual Machine (iVM) technology, iDA provides tools that ensure long-term preservation of databases on physical storage media. This includes not only safeguarding the stored content but also enabling its regeneration with functional search capabilities. For this purpose, two innovative elements are introduced: DbSpec, a new language for managing the decommissioning process, and a Read-Only Access Engine (ROAE) which serves as an interface to future users who wish to retrieve the decommissioned information stored on the long-term substrate. ROAE complements the SIARD (Software Independent Archiving of Relational Databases) standard. Although SIARD is effective at preserving database data and metadata, it lacks the ability to capture the essential search and query functions necessary for meaningful information retrieval. iDA addresses this limitation, ensuring that decommissioned systems remain accessible and functional for future users.

Dette notatet oppsummerer arbeidet gjort i ReDUGUSON-prosjektet. Prosjektet ble gjennomført i samarbeid med VoiceOn AS og finansiert av FORREGION VIKEN. Prosjektet hadde til hensikt å kartlegge hvem som kunne dra nytte av innholdsopplesningstjenester. Vi holdt workshopper med personer med nedsatt syn og personer som hadde hatt slag og hadde afasi. I workshopene diskuterte man temaer knyttet til informasjon på offentlige sider, hjelpemidler og lovgivning. Deltakerne fikk mulighet til å undersøke tjenester som tilbød innholdsopplesning og ga tilbakemeldinger på dem. Noen ideer til forbedringer rundt etikk er presentert her. I tillegg skapte deltakerne sammen utkast til personaer og brukerreiser. De endelige utgavene av disse er presentert i rapporten. Deltakerne hadde også mulighet til å fylle ut et spørreskjema. Spørreskjemaet viste stor nytte av innholdsopplesning for deltakerne. Til slutt presenterer vi noen mulige steder for et hovedprosjekt.

Offentlige virksomheter er i ferd med å ta i bruk generativ kunstig intelligens (KI) for mange ulike oppgaver, blant annet tekstredigering og svar på spørsmål om tekster. For offentlig sektor er regelverk en særlig viktig type tekster, og to sentrale oppgaver knyttet til regelverk, som lover og forskrifter, er utforming og anvendelse av regelverket. Lover og forskrifter følger en bestemt logikk, der teksten er bygget opp av nøstede strukturer av forbud, tillatelser og unntak. For å arbeide med en slik tekst er det nødvendig å forstå hvordan denne strukturen gir den enkelte bestemmelse en semantisk kontekst. Vi studerer i hvilken grad og med hvilken kvalitet generativ KI kan håndtere oppgaver knyttet til forenkling og anvendelser av slikt regelverk. Våre undersøkelser viser at generativ KI per i dag er langt fra å være moden for slike oppgaver.

We propose a Long Short-Term Memory (LSTM) network to estimate historical daily streamflow and hydropower generation in Norway, with particular focus on run-of-river (ROR) plants. Historical records from such plants are often limited, and typically only contain hydropower generation data, which are truncated at the plants’ capacity limits and therefore do not capture high-flow conditions. The proposed LSTM model improves predictions in data-sparse and ungauged catchments, and for high-flow conditions, by learning from both hydropower generation data from ROR plants and streamflow data from other Norwegian catchments. Our model builds upon the neuralhydrology package, by adding a component that transforms streamflow into hydropower generation before loss calculations. The model is trained using streamflow and hydropower generation data from 190 Norwegian catchments and 136 ROR plants, with precipitation, temperature and catchment attributes as input variables. The LSTM model outperforms more traditional hydrological models for predictions in both gauged and ungauged catchments. Furthermore, the combined LSTM model yields hydropower generation estimates that are comparable to or better than those from a model trained only on hydropower generation data, while producing considerably better streamflow estimates. Our approach highlights the added value of additional data sources for hydrological modeling for both local calibration and the task of regionalization, and demonstrates that data-driven methods are suitable for leveraging their potential.

BRACE - et praktisk, datadrevet rammeverk for samhandling om deling av ressurser ved store og samtidige hendelser.

I denne rapporten beskrives arbeidet og resultater fra en utvidelse av validering og evaluering av produkter for snødekningsgrad (FSC) fra 2024. Hensikten har vært å undersøke hvordan avvikene (feil) i FSC-verdier, relativt til «fasiter» fra bilder med høyere oppløsning, varierer i tid og rom med aggregeringsnivå, arealdekke og terreng. Analysene er gjort for Østlandet og tilsigsområder valgt ut av NVE med satellittdata over flere år. Referansedata («fasiter») er basert på Sentinel-2 MSIprodukter i 10 m oppløsning, mens FSC-produktene som ble analysert, er basert på 0,5 km data fra Sentinel-3 SLSTR. Resultatene viser at avvikene har tydelig romlig og tidsmessig struktur. Romlig aggregering reduserer MAE og RMSE, mens bias i hovedsak bevares. En betydelig del av feilen midles likevel ikke effektivt ut ved aggregering opp til de største testede skalaene, noe som tyder på systematiske avvik over større områder. I høydesoneanalysene fremkommer et robust mønster med økte avvik rundt 400-500 m. Når FSC aggregeres til sone-middelverdier per dato reduseres avvikene sammenliknet med pikselbasert stratifisering, men mønsteret i høyde avtar ikke fullt ut. Avvikene er generelt større i skog enn i områder med bart fjell og sparsom vegetasjon, med en negativ bias i skog, som er konsistent med utfordringer knyttet til snø under trekroner. Samlet viser analysen hvordan avvik endrer karakter ved aggregering til modelleringsrelevante enheter, og peker på forhold som bør tas hensyn til ved bruk av FSC som areal- og høydesoneaggregert modellinput

In this study we have investigated how vision foundation models, pretrained on different domains, compete with a specialized model for classification as a function of the size of the labeled training set of medical images. Furthermore, we have looked into the different models' ability to generalize to difficult cases. Our experiments are conducted for cardiac ultrasound images and the downstream task of view recognition. Still, this classification task is meant to serve as a demonstrative example, where we think that the findings should be transferable to other classification tasks and other domains. Through these experiments we found that the foundation models were able to beat the performance of our task-specific supervised model when labelled training data were limited. This was true even for models trained on natural images and when using the simple linear probing method to create a classifier. We observed that more domain-specific foundation models achieved an even higher performance with limited data. On the other hand, the more general models showed a greater ability to generalize and perform well on difficult, out-of-distribution cases. Still, for typical in-domain cases with sufficient labeled data, a task-specific ResNet model was competitive with the foundation models, while also being both smaller and faster.

Wide-beam or single-transmit acquisitions often reduce local spatial coherence, breaking the narrowband model assumed by high-resolution array spectral estimators such as IAA and Capon and thereby degrading performance. We propose a discrete prolate spheroidal sequence (DPSS) subspace projection of delay-focused aperture data. This projection suppresses incoherent off-angle energy and restores local spatial coherence, enabling scanline-wise adaptive spectral estimation under severe model mismatch. Each delay-focused aperture vector is projected onto a DPSS subspace spanned by the first K eigenvectors corresponding to a small angular bandwidth. The approach is lightweight, with precomputation and a per-point complexity of O(MK), and integrates naturally into standard delay-focused processing pipelines. Frequency–angle plots reveal how the projection reconstructs coherent ridge structures that are otherwise obscured by wide‑beam incoherence. Simulations in both plane-wave and diverging-wave ultrasound scenarios demonstrate improved resolution and contrast in single-transmit wide-beam imaging. Qualitative results on recorded channel data from the public PICMUS dataset provide an experimental sanity check and validation, indicating that the same coherence-restoration behavior is observed in real recordings. All experimental validation in this work is confined to ultrasound imaging; assessment of other array-processing applications is left for future work.

Vulnerability databases, such as the National Vulnerability Database (NVD), offer detailed descriptions of Common Vulnerabilities and Exposures (CVEs), but often lack information on their real-world impact, such as the tactics, techniques, and procedures (TTPs) that adversaries may use to exploit the vulnerability. However, manually linking CVEs to their corresponding TTPs is a challenging and time-consuming task, and the high volume of new vulnerabilities published annually makes automated support desirable.This paper introduces Triage, a two-pronged automated approach that uses Large Language Models (LLMs) to map CVEs to relevant techniques from the Att&ck knowledge base. We first prompt an LLM with instructions based on MITRE’s CVE Mapping Methodology to predict an initial list of techniques. This list is then combined with the results from a second LLM-based module that uses in-context learning to map a CVE to relevant techniques. This hybrid approach strategically combines rule-based reasoning with data-driven inference. Our evaluation reveals that in-context learning outperforms the individual mapping methods, and the hybrid approach improves recall of exploitation techniques. We also find that GPT-4o-mini performs better than Llama3.3-70B on this task. Overall, our results show that LLMs can be used to automatically predict the impact of cybersecurity vulnerabilities and Triage makes the process of mapping CVEs to Att&ck more efficient.

Noen mennesker kan ikke kommunisere ved hjelp av talespråk, og vil ha behov for alternativ eller supplerende kommunikasjon (ASK). En av de vanligste formene for ASK er grafiske symboler som står for et ord eller begrep, og brukes av både barn og voksne som deres språklige uttrykksform. Det finnes hverken en gratis symbolbank eller en offisiell standard for grafiske symboler i Norge. Symbolbankene som brukes per i dag krever betaling av lisenser og programmer. Det finnes land som har en åpen, tilgjengelig symbolbank, som Spania (ARASAAC) og Sverige (Bildstöd). Det er lite kunnskap om hva som må løses for å integrere eller utvikle en tilsvarende løsning i Norge. Prosjektet «Demokratisk tilgang til ASK-symboler i Norge» har undersøkt hvordan tilgang til symbolbank oppleves, og om det er behov for en gratis, åpen og universell symbolbank tilgjengelig for alle i Norge. Spørreskjema på nett og to samskapingsverksted med et strategisk utvalg bestående av fagpersoner, ASK-språklige og deres nærpersoner har vært benyttet som metode. På det første samskapingsverkstedet fikk informantene i utvalget beskrive sine erfaringer med tilgang til symbolspråk. Videre ble også et spørreskjema utarbeidet i samarbeid med utvalget. Spørreskjemaet ble sendt ut via nettverk og digitale portaler til hele landet. Forskerne har, sammen med deltakerne i det andre samskapingsverkstedet, systematisert, analysert og dokumentert kunnskapen og erfaringene fra spørreskjemaet. Resultatene fra både spørreskjema og samskapingsverkstedene, viser at tilgang til grafiske symboler oppleves tungvindt og lite tilgjengelig. Det etterspørres en åpen og gratis tilgang for å endre tilgjengelighet, kunnskap om og bruk av et grafisk symbolspråk i Norge. Denne rapporten dokumenterer dette arbeidet. Avslutningsvis er det foreslått et videre arbeid for å nå ambisjonen om en åpen, nasjonal symbolbank i Norge med bakgrunn i studiens funn.

Det ble gitt en oversikt over digitale barrierer som påvirket arbeidsdeltakelsen for personer med funksjonsnedsettelser, og statistikk som viser at mange opplever utfordringer med bruk av IKT‑løsninger, og at digital eksklusjon berører en betydelig del av befolkningen. Videre ble forskjellen mellom universell utforming av IKT og individuelle hjelpemiddelløsninger gjennomgått, samt utfordringene som oppstår når arbeidslivet hovedsakelig bygget på individuell tilrettelegging. Det ble vist til en studie i IDA-prosjektet som tyder på at universelt utformede systemer gir gevinster for både arbeidstakere og virksomheter. Avslutningsvis ble det pekt på at for for å oppnå bedre inkludering i det digitale arbeidslivet er det behov for bedre tilgang til hjelpemidler, tilpasset opplæring og tydeligere krav til universell utforming.

- Vi må frambringe den forskningsbaserte kunnskapen som kan gi oss et bedre grunnlag for å forstå hvordan KI faktisk påvirker læring og vurdering i høyere utdanning, skriver syv ansatte ved en rad NTNU-institutter.

Current benchmarks for long-context reasoning in Large Language Models (LLMs) often blur critical factors like intrinsic task complexity, distractor interference, and task length. To enable more precise failure analysis, we introduce CogniLoad, a novel synthetic benchmark grounded in Cognitive Load Theory (CLT). CogniLoad generates natural-language logic puzzles with independently tunable parameters that reflect CLT's core dimensions: intrinsic difficulty ($d$) controls intrinsic load; distractor-to-signal ratio ($\rho$) regulates extraneous load; and task length ($N$) serves as an operational proxy for conditions demanding germane load. Evaluating 22 SotA reasoning LLMs, CogniLoad reveals distinct performance sensitivities, identifying task length as a dominant constraint and uncovering varied tolerances to intrinsic complexity and U-shaped responses to distractor ratios. By offering systematic, factorial control over these cognitive load dimensions, CogniLoad provides a reproducible, scalable, and diagnostically rich tool for dissecting LLM reasoning limitations and guiding future model development.

We describe different approaches that were used to test and investigate potential improvements in various aspects of the posterior Markov chain. This includes different ways of combining Up and Down chains, and inclusion of all available transitions within a window. Additionally, an alternative algorithm for computing the posterior Markov chain was tested, relying on forward-backward algorithm, using a higher order chain to obtain the posterior Lfcs. Proposed approaches were tested on both PCube paper example and Volund dataset, and results from the latter are presented. First approaches yielded some changes, but not significant improvements implying that current approach is a robust and reasonable choice. On the other hand, the proposed alternative algorithm produced different results. However, it is not straightforward to conclude, whether the results are to be preferred over the ones produced by the existing approach.

We demonstrate that hyperbolic representation learning effectively captures hierarchical cellular relationships in breast cancer. Using information-theoretic metrics, Lorentzian embeddings are shown to preserve significantly more biologically meaningful structure than Euclidean ones. Code: https://github.com/youssefwally/FlatlandandBeyond.

Norge står overfor en demografisk endring med en raskt voksende eldre befolkning. Dette øker presset på helse- og omsorgstjenestene. Samtidig ønsker de fleste eldre å bo trygt og selvstendig i eget hjem så lenge som mulig. Forprosjektet "Smart trygghet for eldre" har tatt utgangspunkt i dette behovet og gjennomført en Proof of Concept (PoC) for en innovativ sensorløsning utviklet av Eldurai AS. Hovedmålet har vært å validere et teknisk konsept som, ved hjelp av sensorer og kunstig intelligens (KI), kan monitorere bevegelsesmønstre og proaktivt avdekke unormale hendelser eller tidlige tegn på forverret helsetilstand. Løsningen er installert og testet i 7 leiligheter hos enslige eldre over 65 år. I samarbeid med Norsk Regnesentral (NR) har prosjektet innhentet brukerinnsikt fra både beboere og pårørende gjennom intervjuer og observasjon. Sentrale forskningsområder er brukeropplevelse, personvern, nytteverdi og balansen mellom trygghet og personlig integritet. Resultatene fra PoC-en, inkludert brukererfaringer, brukskvalitet, nytteopplevelse, aksept, teknisk validering, analyse av bevegelsesmønstre, bidrar til å danne grunnlag for beslutning om videreføring til et hovedprosjekt og en kommersiell løsning

Dette er en oppdatering av notat ADMIN/1/19 der Erik Vasaasen var medforfatter. Hensikten med dette dokumentet er å dokumentere rutinene for behandling av sensitive personopplysninger i NRs forskningsprosjekter i tråd med NRs rutiner for internkontroll. Det sentrale formålet er å fastlegge hvem som er ansvarlig for hva når personopplysninger skal håndteres. Dokumentet inneholder det som kreves etter Personopplysningsloven, med siste oppdatering i juni 2021 som inkorporerer EUs regelverk GDPR.

In this study we have investigated how vision foundation models, pretrained on different domains, compete with a specialized model for classification as a function of the size of the labeled training set of medical images. Furthermore, we have looked into the different models' ability to generalize to difficult cases. Our experiments are conducted for cardiac ultrasound images and the downstream task of view recognition. Still, this classification task is meant to serve as a demonstrative example, where we think that the findings should be transferable to other classification tasks and other domains. Through these experiments we found that the foundation models were able to beat the performance of our task-specific supervised model when labelled training data were limited. This was true even for models trained on natural images and when using the simple linear probing method to create a classifier. We observed that more domain-specific foundation models achieved an even higher performance with limited data. On the other hand, the more general models showed a greater ability to generalize and perform well on difficult, out-of-distribution cases. Still, for typical in-domain cases with sufficient labeled data, a task-specific ResNet model was competitive with the foundation models, while also being both smaller and faster.

Oppsummering av arbeidet i prosjektet «Regelverk,digitalisering og KI» i 2025.