Publications

A fundamental barrier to the adoption of AI systems in clinical practice is the insufficient transparency of AI decision-making. The field of Explainable Artificial Intelligence (XAI) seeks to provide human-interpretable explanations for a given AI model. The recently proposed Neural Explanation Mask (NEM) framework is the first XAI method to explain learned representations with high accuracy at real-time speed. NEM transforms a given differentiable model into a self-explaining system by augmenting it with a neural network-based explanation module. This module is trained in an unsupervised manner to output occlusion-based explanations for the original model. However, the current framework does not consider labels associated with the inputs. This makes it unsuitable for many important tasks in the medical domain that require explanations specific to particular output dimensions, such as pathology discovery, disease severity regression, and multi-label data classification. In this work, we address this issue by introducing a loss function for training explanation modules incorporating labels. It steers explanations toward target labels alongside an integrated smoothing operator, which reduces artifacts in the explanation masks. We validate the resulting Neural Explanation Masks with target labels (NEMt) framework on public databases of lung radiographs and skin images. The obtained results are superior to the state-of-the-art XAI methods in terms of explanation relevancy mass, complexity, and sparseness. Moreover, the explanation generation is several hundred times faster, allowing for real-time clinical applications. The code is publicly available at https://github.com/baerminator/NEM_T

The uptake of practices labeled as "agile" is a topic of widespread discussion in research and practitioner communities. Within the broad topic of agility, there are discussions about the separation between agility in what one might call traditional software development, agility in the form of product orientation, and agility as expressed in continuous delivery. Although particular cases have been studied, the magnitude and manner of adoption and use of agile practices under these three themes remain unclear. We therefore sought to test hypotheses about their growth, prevalence, and implementation patterns in the Norwegian public sector. Aiming to form a comprehensive picture, we distributed surveys three years apart to IT executives at all Norwegian public institutions likely to build digital solutions. The results supported the view that agile practices are prevalent, but gave mixed support for their increase in use. We found no support for the view that agility in, respectively, software development, product orientation, and continuous delivery are treated as distinct disciplines in practice. We were also unable to identify other patterns in implementing these practices. The adoption of agile appears to be enabled primarily by commitment at the team and individual levels and inhibited by factors specific to the public sector. These findings should be compared with other sectors and countries. We propose issues for (a) further research on the scope of agile practices, (b) better indicators for adoption, (c) interaction among agile practices, and (d) factors that enable or inhibit the adoption of agile practices.

The Visual Intelligence Annual Report 2024 highlights the centre's progress, activities and achieved innovations for 2022. It describes new deep learning methods which address pressing societal needs in the fields of medicine and health, marine science, the energy sector, and earth observation.

The Visual Intelligence Annual Report 2021 highlights the centre's progress, activities and achieved innovations for 2021. It describes new deep learning methods which address pressing societal needs in the fields of medicine and health, marine science, the energy sector, and earth observation.

The Visual Intelligence Annual Report 2020 highlights the centre's progress, activities and achieved innovations for 2020. It describes new deep learning methods which address pressing societal needs in the fields of medicine and health, marine science, the energy sector, and earth observation.

Federated learning enables collaborative computation across multiple decentralized devices, minimizing data transfer overhead while enhancing privacy by keeping data local. However, it remains susceptible to inference attacks and potential data leakage. To strengthen privacy guarantees, especially for sensitive domains, advanced privacy-preserving techniques such as homomorphic encryption are recommended. This work proposes a privacy-preserving federated learning framework that integrates threshold homomorphic encryption into the federated learning pipeline to enable secure aggregation and protect intermediate computations. We employ threshold homomorphic encryption, a cryptographic technique well-suited for multiuser environments such as federated learning. We utilize the Cheon-Kim-Kim-Song (CKKS) scheme, as implemented in the OpenFHE library. Our approach extends the standard Federated Averaging (FedAvg) algorithm by homomorphically encrypting model updates and performing aggregation directly on encrypted data. To assess the trade-offs between efficiency and security, we evaluate the performance of the proposed method against a baseline. The design prioritizes practical constraints, including computational efficiency, making it suitable for deployment in privacy-sensitive domains such as healthcare and finance. To ensure compatibility with continuous integration and deployment (CI/CD) pipelines, all components of the solution are containerized using Docker.

Recognizing Urdu text in natural images is more challenging as compared to other languages, such as English, due to the cursive nature of Urdu script. However, Urdu scene text has not received enough attention from both industry and academia due to the lack of the dataset of Urdu text. We propose a large-scale Urdu Scene Text Dataset (USTD) to address this problem, which is designed for Urdu scene text detection and recognition. The proposed dataset contains 29674 text annotations (17877 Urdu and 11797 English), 749725 characters in 6389 images. It covers a wide variety of text images with both Nastaleeq and Naskh writing styles, taken from different streets and roads of Pakistan. The vast diversity of this dataset makes it a benchmark to work on and train robust neural networks for the detection and recognition of cursive text. Besides, baseline results are also provided with several state-of-the-art networks, including TextBoxes++, Seglink, DB(ResNet-50) and EAST for text localization and Convolutional Recurrent Neural Network (CRNN) for text recognition. To further evaluate the performance of these models, we have used the most popular evaluation matrices of precision, recall, and F-measure. Our experimental outputs reveal that an end-to-end combination of DB(ResNet-50) and CRNN provides the best results with precision, recall, and F-measure of 0.7526, 0.5974, and 0.6660, respectively.

Dette er dokumentasjon av programvaren i LAVDAS slik den foreligger per mars 2025

In the AIforScreening project we have tested different ways of utilizing time sequence information in mammography screening, including published methods and home-grown ones. The simplest ones include regression modelling, where we apply a single-image breast cancer risk model at a sequence of images of a given breast and use linear regression to onstruct a modified score. This method givs a very modest improvement of the order 0.001 on the AUC scale, which was statistically significant only for the inferior holistic model. The more advanced methods try include co-registration of the current and previous images and various ways of merging the model’s features to produce improved risk scores, utilizing various so-called Siamese net models. Over-all, the results were negative, as none of the advanced methods gave improvements above the linear model. This is contrary to published results, and we speculate that this may be due to the fact that our model has a high performance to begin with, leaving less room for improvement. The linear model places positive weight on the previous risk scores, which go against the intuition that an increase in risk score over time should increase the likelihood of cancer. Apparently, the ’direct effect’ that an elevated risk score is associated with future cancer is stronger.

Health systems in Norway and across Europe are under increasing strain from chronic conditions, long waiting times, and limited clinical capacity. At the same time, evidence-based digital therapeutics (DTx) are emerging as regulated tools to prevent, manage, and treat disease through clinically validated software interventions. Scandinavia has played a pioneering role in evaluating internet-delivered cognitive behavioural therapy for conditions such as insomnia and perinatal depression, demonstrating that well-designed, integrated digital solutions can complement existing healthcare services. However, despite promising trial evidence, large-scale adoption of DTx remains inconsistent due to regulatory, reimbursement, organisational, and cultural barriers. In parallel, simulation-based methods such as discrete-event modelling and digital twins are increasingly applied to optimise healthcare delivery and test alternative service configurations. Building on these developments, this project proposes the creation of an adaptive digital twin of a chronic care pathway to analyse how different deployment strategies for prescription digital therapeutics could impact system access, resilience, and resource utilisation.

In today’s rapidly evolving geopolitical climate, cybersecurity and digital sovereignty are more critical than ever for Europe’s resilience—and for investments in AI, healthcare and other critical sectors.

This user manual describes the COHIBA surface modeling software. It consists of: Part I Introduction: Basic ideas and terminology Part II User manual: Usage, input data, and results Part III Tutorials: Special topics such as volumes, simulation, and faults Part IV Reference manual: Descriptions of all COHIBA model file elements Part V Theory: Methods used by COHIBA Part VI Appendix: Release notes, known issues, references, list of acronyms, tables and figures, and an index

This user manual describes the COHIBA surface modeling software. It consists of: Part I Introduction: Basic ideas and terminology Part II User manual: Usage, input data, and results Part III Tutorials: Special topics such as volumes, simulation, and faults Part IV Reference manual: Descriptions of all COHIBA model file elements Part V Theory: Methods used by COHIBA Part VI Appendix: Release notes, known issues, references, list of acronyms, tables and figures, and an inde

Time series data is fundamentally important for many critical domains such as healthcare, finance, and climate, where explainable models are necessary for safe automated decision-making. To develop explainable artificial intelligence in these do mains therefore implies explaining salient information in the time series. Current methods for obtaining saliency maps assumes localized information in the raw input space. In this paper, we argue that the salient information of a number of time series is more likely to be localized in the frequency domain. We propose FreqRISE, which uses masking-based methods to produce explanations in the frequency and time-frequency domain, and outperforms strong baselines across a number of tasks. The source code is available here: https://github.com/theabrusch/FreqRISE.

Two new different stochastic models for earthquake occurrence are discussed. Both models are focusing on the spatio-temporal interactions between earthquakes. The parameters of the models are estimated from a Bayesian updating of priors, using empirical data to derive posterior distributions. The first model is a marked point process model in which each earthquake is represented by its magnitude and coordinates in space and time. This model incorporates the occurrence of aftershocks as well as the build-up and subsequent release of strain. The second model is a hierarchical Bayesian space-time model in which the earthquakes are represented by potentials on a grid. The final ambition of the models is to make predictions on the occurrence of earthquakes.

The asymptotic behavior of several goodness-of-fit statistics for copula families is obtained under contiguous alternatives. Many comparisons between a Craméer-von Mises functional of the empirical copula process and new moment-based goodness-of-fit statistics are made by considering their associated asymptotic local power curves. It is shown that the choice of the estimator for the unknown parameter can have a significant influence on the power of the Craméer-von Mises test, and that some of the moment-based statistics can provide simple and efficient goodness-of-fit methods. The paper ends with an extensive simulation study that aims to extend the conclusions to small and moderate sample sizes.

In this paper we develop statistical models for bankruptcy prediction of Norwegian firms in the limited liability sector using annual balance sheet information. We fit generalized linear-, generalized linear mixed- and generalized additive models in a discrete hazard setting. It is demonstrated that careful examination of the functional relationship between the explanatory variables and the probability of bankruptcy enhances the models' forecasting performance. Using information on the industry sector we model the unobserved heterogeneity between different sectors through an industry-specific random factor in the generalized linear mixed model. The models developed in this paper are shown to outperform the model with Altman's variables at all levels of risk. As a measure of models' forecasting accuracy the area under the ROC curve is used.

Several copula goodness-of-fit approaches are examined, three of which are proposed in this paper. Results are presented from an extensive Monte Carlo study, where we examine the effect of dimension, sample size and strength of dependence on the nominal level and power of the different approaches. While no approach is always the best, some stand out and conclusions and recommendations are made. A novel study of p-value variation due to permuation order, for approaches based on Rosenblatt's transformation is also carried out. Results show significant variation due to permutation order for some of the approaches based on this transform. However, when approaching rejection regions, the additional variation is negligible. Finally, motivated by the permutation study, new versions of some goodness-of-fit approaches are proposed and examined. The new versions consider all permutation orders of the variables and we see some power improvement over the approaches that consider one permutation order only.

Historisk befolkningsregister (HBR) er et autoritetsregister over alle personene i norsk personhistorisk kildemateriale tilbake til 1801 og gjør det mulig å identifisere og gjenfinne personene i disse kildene. Hver person får en unik ID som er viktig for dokumentasjon og for å finne mer informasjon om vedkommende. HBR utvider perioden med detaljert kunnskap om hver person i den norske befolkningen fra 60 år i dagens folkeregister til 224 år, fra to til syv generasjoner. I denne artikkelen beskriver vi oppbyggingen av HBR med lenkingsstrategier samt personvern og hvordan det avviker fra andre befolkningsoversikter. Videre blir det drøftet hvordan man kan oppnå best mulig kvalitet i registeret, samt representativitet og hvor høy lenkingsgrad det er mulig å oppnå. Kunnskap om hver enkelt persons livsløp, bosted og familie gir en ny innsikt i befolkningen og åpner for nye forskningsmetoder og mer presise beskrivelser og analyser. Dette vises ved å gi et datagrunnlag for migrasjon 1910–1920 og en analyse av fødselsdatoer i folketellinger. Forskere vil kunne trekke ut data om de problemstillinger de interesserer seg for, som grunnlag for sin egen forskning.

Relevant factors can have shifting prognostic impacts on cardiovascular disease (CVD) occurrences among in patients with breast cancer (BC) over time. CVD incidence and its driving factors vary among different CVDs. We examined the time to the first occurrence of heart attack, atrial fibrillation, embolic stroke, angina pectoris, embolism, peripheral vascular disease, and heart failure (HF). We particularly focused on the influence of molecular subtype, adjusting for age, tumor stage, and radiation therapy.

Maintaining the performance of a deep learning model trained for breast cancer detection on a specific scanner type is challenging in a multi-scanner setting due to domain shifts caused by variations in imaging data. This often results in a performance drop when models trained on one scanner are tested on another. While re-training with labeled data from the new scanner is an option, delays in obtaining ground-truth labels make this approach impractical. To overcome this limitation, Unsupervised Domain Adaptation (UDA) offers a promising alternative by enabling models to adapt across scanners without requiring labeled target data. In this study, we investigated Conditional Domain-Adversarial Network (CDAN), an adversarial UDA approach, to adapt a classifier trained on Siemens scanner data using nearly 3 million mammograms from the Norwegian breast cancer screening program. We compared it to Maximum Mean Discrepancy (MMD), a simpler statistical feature alignment method, and evaluated histogram matching, which required no additional training. Our findings showed that the AUC drop on the target GE data (0.96 to 0.62) without adaptation was mitigated by histogram matching (AUC 0.84), but that was less effective than MMD (AUC 0.87), which performed competitively with CDAN. Further ablation with Domain-Adversarial Neural Network (DANN), the foundation of CDAN, suggested limitations in the domain discriminator. Unlike prior work focusing solely on performance, we paired UDA with explainability. This revealed how feature relevance shifted across scanner domains, offering novel insights into model generalizability in cancer detection.

A linear regression model is developed to link anomalies of streamflow to anomalies of precipitation amounts and temperature with the goal of making multi-decadal streamflow projections based on CMIP6 multi-model simulations. Regression coefficients estimated separately for each catchment and each month show physically implausible spatial patterns and indicate issues with overfitting. An alternative approach is therefore explored in which all regression coefficients are estimated simultaneously through a neural network that retains the original linear model structure, but uses embeddings to map each combination of catchment and month to a set of regression coefficients. The model is demonstrated over a set of catchments in Brazil, where the estimated relationships are used to make streamflow projections for the next decades based on CMIP6 multi-model simulations. It yields physically more plausible relationships between streamflow, precipitation amounts, and temperature for our study area than the locally fitted regression models. The resulting projections indicate reduced streamflow over northern, north-eastern, central, and south-eastern Brazil, especially for the austral spring and summer season. The signal is less clear during austral winter. In southern Brazil, an increase in streamflow is expected.

Advice on fishing opportunities is based on stock assessment models. Modern stock assessment frameworks are designed to account for uncertainty in model parameters. However, some parameters are often fixed at specific values and assumed to be known precisely. Paradoxically, these fixed values often relate to parameters we understand the least—such as the natural mortality rate. For example, in the assessment of Norwegian Spring-Spawning Herring, the natural mortality rate is assumed to be exactly 0.9 for age 2 fish and 0.15 for older fish. In this study, we investigate how sensitive the assessment results—and, more specifically, the resulting fishing quotas (total allowable catch, TAC)—are to assumptions regarding natural mortality and maturity-at-age for both Norwegian Spring-Spawning Herring and North-East Arctic Cod. The assessment models for both stocks are formulated within the SAM framework. Furthermore, we explore a simple approach for adjusting harvest control rules that seeks to preserve the TACs when fundamental fixed assumptions are altered. Our intention is that this serves as an exploratory tool to investigate sensitivity to new assumptions, and is not intended to replace a more formal process for defining new biological reference points and harvest control rules. The results indicate that, under the current harvest control rules, the TACs for both herring and cod may be sensitive to assumptions regarding natural mortality rates and maturity-at-age. For herring, sensitivity to natural mortality is significantly greater than is reflected by the estimation uncertainty in model parameters. However, when basic assumptions are changed, it is also natural to consider corresponding adjustments to the harvest control rules. Our proposed simple adjustments to the control rules reduce sensitivity to these assumptions in many cases, though not in all.

The main objective of the working group was to review recent surveys of Greenland Sea harp and hooded seal pup production and examine harvest scenarios for these populations as well as harp seals in the White Sea. No new survey to estimate pup production of Barents Sea/White Sea harp seals was completed. No new survey information was available for the Northwest Atlantic. The 2022 Greenland Sea aerial survey images were analyzed manually and with the aid of automatic detection methodology (deep learning). For assessment purposes, this report only refers to the manual counts. Correction factors based on staging surveys were applied according to established methodology. The 2022 Greenland Sea harp seal pup production estimate for harp seals was 92,769 (CV = 20.2%), which is significantly higher than the 2018 estimate but similar to that based on the 2012 survey. The hooded seal pup production estimate for 2022 was 13,509 (CV=12.9%), slightly but not significantly higher than the 2018 estimate. Subsequent to the recent benchmark meeting, model development indicated that the model estimates of adult population size for the Greenland Sea population of harp seals is highly sensitive to the standard deviation on the prior for initial population size. The WG therefore concluded that the current version of the assessment model could not be used to explore harvest scenarios based on estimates of current or projected total population size. Moreover, given the fact that the estimate of current total population size is unreliable, it also did not allow for robust calculation of Potential Biological removals (PBR). Tentatively, two different approaches are presented that might be used to inform sustainable harvest levels until the model has been further improved and reviewed: 1) an adaptive management approach based on population trends and 2) PBR based on a conservative population estimate that is a simple scaling of the observed levels of pup production, based on plausible values of adult:pup ratios. The Greenland Sea hooded seal population shows continued decline, and remains below the Lower Reference Limit despite no hunting since 2007. In a recent review of the status of the Northwest Atlantic harp seal population, model fit to aerial survey estimates of pup production and annual reproductive rates was poor compared to previous assessments indicating underlying problems relating to model assumptions and/or structure. A new hierarchical Bayesian state-space model was fitted to the same data on pup production, annual fecundity, human removals, and environmental conditions used in the previous assessment to produce annual estimates of pup production and total abundance from 1952 - 2019. Data on age structure based upon random samples were also included, and the process model incorporated environmental stochasticity and several other improvements. The new model estimates were similar to the previous model through 1990 but then diverged, indicating that the population peaked in 1997 at 6.6 million animals, almost a decade earlier than modelled in previous assessments. After a period of decline due to high catches and poor ice conditions, the new model provides an abundance estimate of 4.7 (95% Credibility Interval (CI) 3.7-5.7 ) million in 2019, compared to an estimate of 7.6 (95% CI 6.6-8.8) million in the last assessment. The lower estimates of recent abundance reflect higher and more variable juvenile mortality after 2000 due to a combination of density-dependent and density-independent factors operating on juvenile survival. The new model also suggests a decline in equilibrium abundance (K) levels from 7.6 (95% CI=7.4 to 7.8) million Northwest Atlantic harp seals prior to 2000 to 6.8 (95% CI=6.7 to 6.9) million animals post-2000.

The benchmark was tasked with evaluating proposed developments to the assessment model used for two stocks of harp seals (East Ice [White Sea/Barents Sea, seh.27.1] and West Ice [Greenland, seh.27.125a14]) and one stock of hooded seals (West Ice [Greenland, sez.27.2514]) in the Northeast Atlantic. The benchmark concluded that there were sufficient data to produce an assessment model for the West Ice (Greenland Sea) stock of harp seals but that data were insufficient for the East Ice (Barents Sea / White Sea) harp seal stock and too weak a signal for the West Ice hooded seals for viable assessments for these stocks.

Tracer kinetic modeling serves a vital role in diagnosis, treatment planning, tracer development and oncology, but burdens practitioners with complex and invasive arterial input function estimation (AIF). We adopt a physics-informed CycleGAN showing promise in DCE-MRI quantification to dynamic PET quantification. Our experiments demonstrate sound AIF predictions and parameter maps closely resembling the reference.

Tracer kinetic modeling serves a vital role in diagnosis, treatment planning, tracer development and oncology, but burdens practitioners with complex and invasive arterial input function estimation (AIF). We adopt a physics-informed CycleGAN showing promise in DCE-MRI quantification to dynamic PET quantification. Our experiments demonstrate sound AIF predictions and parameter maps closely resembling the reference.

A spatial model implemented in TMB was presented (mslt). The model was recently published in JASA (https://doi.org/10.1080/01621459.2025.2566422) . It treats whale sightings as arising from a thinned Cox process, with intensity driven by a latent Gaussian Markov random field (using the SPDE approach) and a two-state Markov-modulated Poisson process (MMPP). The SPDE component is intended to capture long-range spatial structures, wheras the MMPP accounts for short-range structure. It was demonstrated that including the MMPP component typically increases the spatial correlation length of the SPDE component. This implies that the model can propagate the SPDE process further away from the transects when the MMPP is successfully included. The model jointly estimates the detection function, group intensity, and group size, ensuring that uncertainty is propagated automatically through the delta method implemented in TMB.

Hyperbolic representation learning has shown compelling advantages over conventional Euclidean representation learning in modeling 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 composition 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.

Hyperbolic representation learning has shown compelling advantages over conventional Euclidean representation learning in modeling 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 composition 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 neighbor estimators to rigorously quantify the clustering performance in each geometry using mutual information and conditional mutual information. Our results reveal that hyperbolic embeddings retain significantly more biologically relevant information than their Euclidean counterparts. 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 hyperbolic 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

Hyperbolic representation learning has recently emerged as a powerful framework for modeling hierarchical structures in data, often outperforming Euclidean embeddings. We investigate its utility for analyzing high-dimensional biological data from Imaging Mass Cytometry (IMC) of breast cancer tissues. We embed cells into Euclidean and Lorentzian latent spaces via a fully hyperbolic variational autoencoder (VAE) and propose an information-theoretic framework based on k-nearest neighbor estimators to quantify clustering quality using mutual information (MI) and conditional mutual information (CMI). Results show that Lorentzian embeddings preserve substantially more biologically relevant structure compared to Euclidean ones. We further provide open-source tools for Lorentzian MI estimation and hyperbolic UMAP visualization, enabling geometry-aware representation learning for spatial biology.

Machine learning and statistical methods can improve conventional motor protection systems by providing early warning and detection of emerging failures. Data-driven methods rely on historical data to learn how the system is expected to behave under normal circumstances. An unexpected change in the underlying system may alter the statistical properties of the data, thereby affecting the performance of the fault detection algorithm in terms of time to detection and false alarms. This kind of change, called concept drift, requires adaptations to maintain constant performance. In this article, we present a machine learning approach for detecting overheating in the stator windings of marine electrical propulsion motors. Using simulated overheating faults injected into operational data, the methods are shown to provide early detection compared to conventional methods based on temperature readings and fixed limits. The proposed monitors are designed to operate for a type of concept drift observed in operational data collected from a specific class of motors in a fleet of ships. Using a mix of real and simulated concept drifts, it is shown that the proposed monitors are able to provide early detections during and after concept drifts, without the need for full model retraining.

Efforts towards endowing robots with the ability to speak have benefited from recent advancements in natural language processing, in particular large language models. However, current language models are not fully incremental, as their processing is inherently monotonic and thus lack the ability to revise their interpretations or output in light of newer observations. This monotonicity has important implications for the development of dialogue systems for human–robot interaction. In this paper, we review the literature on interactive systems that operate incrementally (i.e., at the word level or below it). We motivate the need for incremental systems, survey incremental modeling of important aspects of dialogue like speech recognition and language generation. Primary focus is on the part of the system that makes decisions, known as the dialogue manager. We find that there is very little research on incremental dialogue management, offer some requirements for practical incremental dialogue management, and implications of incremental dialogue for embodied, robotic platforms in the age of large language models.

Knowledge graphs are often used to represent structured information in a flexible and efficient manner, but their use in situated dialogue remains under-explored. This paper presents a novel conversational model for human--robot interaction that rests upon a graph-based representation of the dialogue state. The knowledge graph representing the dialogue state is continuously updated with new observations from the robot sensors, including linguistic, situated and multimodal inputs, and is further enriched by other modules, in particular for spatial understanding. The neural conversational model employed to respond to user utterances relies on a simple but effective graph-to-text mechanism that traverses the dialogue state graph and converts the traversals into a natural language form. This conversion of the state graph into text is performed using a set of parameterized functions, and the values for those parameters are optimized based on a small set of Wizard-of-Oz interactions. After this conversion, the text representation of the dialogue state graph is included as part of the prompt of a large language model used to decode the agent response. The proposed approach is empirically evaluated through a user study with a humanoid robot that acts as conversation partner to evaluate the impact of the graph-to-text mechanism on the response generation. After moving a robot along a tour of an indoor environment, participants interacted with the robot using spoken dialogue and evaluated how well the robot was able to answer questions about what the robot observed during the tour. User scores suggest an improvement in the perceived factuality of the robot responses when the graph-to-text approach is employed compared to a baseline using inputs structured as semantic triples.

The finance industry depends on effective anti-money laundering (AML) systems to ensure compliance and maintain operational efficiency. However, existing AML systems, which are predominantly rule-based, frequently struggle to detect money laundering accurately. In particular, their inability to learn from historical data and properly account for diverse customer behavior is problematic. Also accounting for the vast amounts of transactional data generated daily, this challenge calls for big data analytics and advanced machine learning techniques. In line with this, the present paper explores a graph neural network (GNN) approach, a state-ofthe-art machine learning technique, to identify money laundering activities within a large heterogeneous network constructed from real-world bank transactions and business role data from DNB, Norway’s largest bank. To this end, we extend the (homogeneous) Message Passing Neural Network (MPNN) architecture to operate on a heterogeneous graph, and demonstrate its strong performance in detecting money laundering activities. We showcase the suitability of utilizing GNN methodology to improve electronic surveillance systems for detecting money laundering, thereby contributing a pioneering approach to AML through the application of advanced data science techniques. To the best of our knowledge, this is the first publication applying heterogeneous GNNs for AML purposes with a large real-world heterogeneous network.

Glioblastoma is characterized by diffuse infiltration, making accurate detection of residual disease essential for improving prognostication and guiding treatment. This study evaluates whether the volume of predicted infiltration, generated by a machine learning (ML) model trained on radiomic features from postoperative magnetic resonance imaging (MRI), is an independent prognostic factor. We analyzed a total of 114 glioblastoma patients, 89 from a retrospective multicenter cohort and 25 from a prospective cohort, who underwent gross total resection and had an early postoperative MRI. A previously published voxel-wise ML model estimated tumor infiltration probability in the non-enhancing peritumoral region using conventional MRI sequences. High-risk of recurrence regions (HRoR) were delineated from the probability maps, and their volumes were quantified. Associations with residual FLAIR volume, clinical variables (age, Karnofsky Performance Status), and survival outcomes (overall survival [OS], progression-free survival [PFS]) were evaluated using Cox regression and Kaplan–Meier analysis. In the retrospective cohort, multivariate Cox modeling confirmed that higher HRoR volume was independently associated with shorter OS (HR = 1.51; 95% CI, 1.12–2.05; p = 0.008), with no association found for PFS. A robust cutoff of 1.6 cm³ stratified patients into high- and low-risk groups with significantly different OS (456 vs. 678 days; p = 0.038). This threshold was validated in a prospective cohort (326 vs. 525 days; p = 0.039). ML-derived HRoR mapping provides independent prognostic value and may improve risk stratification after surgery in glioblastoma. These findings support its potential clinical integration for personalized follow-up and treatment.

The evolution of 5G networks toward 6G presents an opportunity to shift from centralized security to decentralized, intelligent, and autonomous security. While 5G introduces decentralization through software-based virtualization and edge computing, it also exposes network infrastructure to a new and dynamic threat landscape that still relies largely on static, centrally deployed threat detection. This work proposes a lightweight anomaly detection framework based on Tiny Machine Learning (TinyML), specifically designed to address the latency, memory, and energy constraints of 5G edge devices. Using promising model compression techniques, including integer quantization and pruning, we demonstrate that highly optimized models can run entirely on-device without relying on cloud infrastructure. Our experiments using a real-world 5G dataset demonstrate that the quantized TinyDenseNet achieves over 99% detection accuracy while keeping the model size below 30 KB and inference latency under 11 ms. By embedding intelligent detection at the network edge, this approach presents self-defending, ultra-reliable, and context-aware infrastructures for future 6G networks. The proposed approach focuses on next-generation networks, in which decentralized, low-power, and privacy-preserving security mechanisms will be essential.

Clinical guidelines recommend performing left ventricular (LV) linear measurements in B-mode echocardiographic images at the basal level—typically at the mitral valve leaflet tips—and aligned perpendicular to the LV long axis along a virtual scanline (SL). However, most automated methods estimate landmarks directly from B-mode images for the measurement task, where even small shifts in predicted points along the LV walls can lead to significant measurement errors, reducing their clinical reliability. A recent semi-automatic method, EnLVAM, addresses this limitation by constraining landmark prediction to a clinician-defined SL and training on generated Anatomical Motion Mode (AMM) images to predict LV landmarks along the same. To enable full automation, a contour-aware SL placement approach is proposed in this work, in which the LV contour is estimated using a weakly supervised B-mode landmark detector. SL placement is then performed by inferring the LV long axis and the basal level—mimicking clinical guidelines. Building on this foundation, we introduce WiseLVAM—a novel framework for fully automated yet manually adaptable framework for automatically placing the SL and then automatically performing the LV linear measurements in the AMM mode. WiseLVAM utilizes the structure-awareness from B-mode images and the motion-awareness from AMM mode to enhance robustness and accuracy with the potential to provide a practical solution for the routine clinical application. The source code is publicly available at https://github.com/SFI-Visual-Intelligence/wiselvam.git.

The escalating demand for high-resolution Earth Observation (EO) data for various applications has significantly influenced advancements in image processing techniques. This study proposes a workflow to super-resolve the 12 spectral bands of Sentinel-2 Level-2A imagery to a ground sampling distance of 2.5m. The method leverages a hybrid approach, integrating advanced diffusion models with image fusion techniques. A critical component of the proposed methodology is super-resolution of Sentinel-2 RGB bands to generate a super-resolved Sentinel-2 RGB image, which subsequently serves in the image fusion pipeline that super-resolves the remaining spectral bands. The super-resolution algorithm is based on a diffusion model and is trained using the extensive National Agriculture Imagery Program (NAIP) dataset of aerial images, which is freely available. To make the super-resolution algorithm, trained on NAIP images, applicable to Sentinel-2 imagery, image harmonization and degradation were necessary to compensate for the inherent differences between NAIP and Sentinel-2 imagery. To address this challenge, we utilised a sophisticated degradation and harmonisation model that accurately simulates Sentinel-2 images from NAIP data, ensuring the harmonised NAIP images closely mimic the characteristics of Sentinel-2 observations post-resolution reduction. To investigate if learning the diffusion model using a large dataset of airborne images like NAIP provides better results than learning the model using a smaller satellite-based dataset like WorldStrat of high-resolution SPOT images, we performed a comparative analysis. The results demonstrate that models trained with the harmonised and correctly simulated datasets like NAIP significantly outperform those trained directly on SPOT images but also other existing super-resolution models available. This finding reveals that learning with more data can be beneficial if the data is properly harmonised and degraded to match the Sentinel-2 images. We performed a comprehensive evaluation using the recently established open-SR test methodology to validate the proposed model across multiple super-resolution metrics. This testing framework rigorously evaluates the super-resolution model based on metrics beyond traditional super-resolution metrics such as PSNR, SSIM, and LPIPS. Instead, the open-SR test evaluates the model based on metrics that measure its consistency, synthesis, and correctness. The proposed super-resolution model outperformed several current state-of-the-art models based on the comprehensive open-SR test framework. In addition, visual comparison further established the superior performance of our model in both urban and rural scenarios. An important component of the proposed model is the super-resolution of all 12 Sentinel-2 Level-2A bands, contrary to previous work, which has mainly focused on RGB band super-resolution. The proposed fusion pipeline successfully utilises the super-resolved image to obtain an enhanced 12-band Sentinel 2 image, similar to pansharpening techniques. We show qualitative and quantitative results on all 12 bands that demonstrate the seamless performance of the fusion method in super-resolution. This study not only showcases the potential of combining AI-driven super-resolution models with image fusion techniques in enhancing EO data resolution but also addresses the critical challenges posed by the diversity in data sources and the necessity for accurate generative models in training neural networks for super-resolution tasks.

This note is a summary of the work done in the Standards for Universal Design of Robots Project. It includes results from the standards search, a summary of standards purchased in the project, requirements generated for our context of a robot used in a daycare or special school, and updated requirements for the Norwegian context. We also document how using standards can be combined with the work from accessibility testing in the previous project on universal design of robots.

Stock assessments are often based on age-structured data obtained by interpreting calcified structures. Due to readability and human error, the observed age may be wrong. We propose a parametric model for age-reading error matrices, which is more realistic and robust than the commonly used empirical matrices. The parameters have meaningful interpretations, allowing for direct comparison of age-reading properties. We compare different species (Atlantic mackerel ( Scomber scombrus) and herring ( Clupea harengus)), stocks (North Sea autumn-spawning vs. Norwegian spring-spawning herring), and calcified structures (otoliths vs. scales). Three out of four data sets had an asymmetry tendency towards reading higher ages than the true age. The estimated probability of reading the wrong age was lower for scales than for otoliths. The true age is often unknown and assumed to be the modal age. We assess the systematic bias due to this assumption. Finally, when including age-reading error in stock assessment, the dominating age classes were estimated to be larger and spawning stock biomass lower. Our study contributes with methods and insight for including age-reading error in stock assessment.

Current debate on the sustainability of using cleaner fish to control parasitic lice in salmon farming suffers from extremely variable scientific evidence on the efficacy of this practice. This paper presents novel experimental results on evacuation rates of salmon lice through the digestive tract of ballan wrasse. These results are combined with quantitative field data on contents of salmon lice in ballan wrasse, to derive a method to study the efficacy of ballan wrasse in cleaning salmon of salmon lice. From a fitted binomial regression model on the probability of finding lice in the digestive tract after ingestion, we found a median evacuation time of 11.0 h. The mean evacuation time was 12.2 h. Furthermore, by integration, we found that if a wrasse on average consumes one louse per day, then the expected number of observable lice in the digestive tract is 0.472. This gave an estimated daily consumption of salmon lice per wrasse expressed as the number of salmon lice in the digestive tract divided by 0.472. As an example, analyses of lice contents in the digestive tract of 6406 ballan wrasse used as cleaner fish in salmon farming revealed that salmon lice were found in 2.9% of the wrasses, with a mean number of 0.15 lice per fish. This translates to an estimate of 0.32 lice consumed per day per ballan wrasse. The present way of estimating the efficacy of wrasse as cleaner fish may contribute to a more robust evaluation of louse control effects of ballan wrasse.

The ESORICS 2025 Workshops proceedings deal with up to date research and practical applications in computer security.

En ny studie viser at mye lus på utvandrende laksesmolt har en klar sammenheng med at færre gytefisk vender tilbake etter ett år. Forskere ved AquaLife R&D, Norsk Regnesentral og Veterinærinstituttet har utviklet en statistisk modell for å undersøke sammenhengen. Vi tallfestet i vår modell i hvilken grad lusebelastningen fisken opplevde på sin smoltvandring har påvirket antall og alder på gytefisk i norske elver. Det vi fant var at mye lus på utvandrende laksesmolt har en klar sammenheng med at færre gytefisk vender tilbake etter ett år i sjø. På fisk som hadde vært to år i sjø var den samme effekten ikke til stede. På fiskene som hadde vært tre år i sjø fant vi motsatt effekt; det var en klar sammenheng mellom høy lusebelastning på utvandrende smolt og økt antall fisk som returnerte som gytefisk, men denne økninga var langt mindre enn den tilhørende nedgangen i antall fisk som hadde vært ett år i sjø. Dette vises i figuren under. Én mulig årsak til denne forskjellen i effekt mellom årsklassene er at lusepåkjenning kan forsinke vekst og kjønnsmodning.

Abstract Negative effects of salmon lice from salmon farms on wild salmonids have been a controversial issue for decades and concerns are expressed from virtually all areas where salmon farming coexists with important fisheries for wild salmonids. A key question is to what extent lice infestation from farms reduce numbers of wild mature salmon returning to spawn in the rivers. Here, we study counts and recreational catches of small (<3 kg), medium‐sized (3–7 kg) and large mature salmon (>7 kg) returning to spawn in Norwegian rivers, in association with lice burdens on out‐migrating post‐smolt recruits. The expected number of returning mature salmon was modelled as a function of theoretical smolt production in the rivers, river catches one generation back in time, size‐classes of returning salmon, year as a factor, a non‐linear spatial effect and parasite‐induced mortality (PIM) of out‐migrating post smolts of salmon. PIM was attributed to small, medium‐sized and large salmon assuming they spend one, two or three winters at sea (SW), respectively. There was a significant negative effect of PIM on returns of one SW salmon and a negative but non‐significant effect on returns of two SW salmon. For three SW salmon, the effect of PIM was significantly positive, but for comparably low numbers, implying an overall negative effect of PIM on returning salmon. The size‐specific effects of PIM were manifested by decreasing proportions of one SW salmon in returning populations with increasing PIM, from ~0.6 for rivers exposed to low levels of PIM, to predictions of <0.2 for rivers exposed to high levels of PIM. Synthesis and applications . This study presents a quantitative relationship between infestations of post‐smolt recruits and size‐structured returns of mature salmon to Norwegian rivers, suggesting that louse infestation from farms may reduce returns of spawners and re‐structure the size distribution of mature river populations of Atlantic salmon. The presented relationship opens for a more targeted approach to obtaining sustainable salmon farming. To accommodate the Norwegian Government's goal for sustainable aquaculture, reductions in lice abundances in farms are necessary.

Dette er en webapplikasjon for nettbrett eller PC som gjør det mulig å fjernstyre en NAO-robot. Med løsningen kan man enkelt bestemme hva roboten skal si og gjøre i dialog med en elev. Applikasjonen inneholder en rekke ferdigprogrammerte fraser, danser og bevegelser, men brukeren kan også skrive inn egen tekst som roboten skal si. En versjon av prototypen fungerer med både NAO og Misty II.

Et grensesnitt hvor man lett kan starte fjernkontroll fra ROSA-repertoaret, med fokus på enkel tilgang og effektiv oppstart.

This article concerns hybrid combinations of empirical and parametric likelihood functions. Combining the two allows classical parametric likelihood to be crucially modified via the nonparametric counterpart, making possible model misspecification less problematic. Limit theory for the hybrid likelihood function is sorted out, also outside of the parametric model conditions. We prove a profiling result as well as limiting behaviour of the maximizer of the hybrid likelihood function. Our results allow for the presence of plug-in parameters in the hybrid and empirical likelihood framework. Furthermore, the variance and mean squared error of these estimators are studied, with recipes for their estimation. The latter is used to define a focused information criterion, which can be used to choose how the parametric and empirical part of the hybrid combination should be balanced. This allows for hybrid models to be fitted in a context driven way.

This study investigates the applicability of the Mammo-CLIP Dissect framework from Salahuddin et al. for concept-based explainability in deep learning models for mammography. We first reproduced key results from the original paper using the Mammo-CLIP image encoder and a curated mammography concept vocabulary, confirming the expected layer-wise emergence of clinically meaningful concepts. We then applied the framework to an in-house ResNet101 classifier developed within the AIforscreening project. Compared with Mammo-CLIP, the ResNet101 model exhibited lower semantic alignment and a higher prevalence of non-mammography concepts, reflecting differences in the absence of multimodal training. These findings suggest that models trained solely on images may provide less interpretable explanations for clinicians than multimodal vision-language models. We highlight the importance of jointly considering accuracy and interpretability, noting that model performance was not evaluated on the probe set in this study. Future work includes applying the framework to Cancer Registry data and exploring multimodal training for improved clinical relevance.

Sammendrag Respons Analyse gjennomførte undersøkelsen «Barometer for synslikestiling» for Norges Blindeforbund i 2023. Resultatene fra undersøkelsen er beskrevet i rapporten “Norges Blindeforbund: Likestillingsbarometeret 2023.” Forskningsprosjektet i 2025 bygger på dette arbeidet. Denne rapporten presenterer resultatene fra forskningsprosjektet om videreutvikling av barometer for synslikestilling, i det følgende kalt Synsbarometeret. Forskningsprosjektets mål har vært å undersøke hvordan styrke kvaliteten og representativiteten til Synsbarometeret, slik at det kan brukes som en pålitelig kilde til informasjon for myndigheter og fagmiljøer. Arbeidet har omfattet en gjennomgang av det tidligere spørreskjemaet og forslag til endringer for den nye telefon-undersøkelsen som ble gjennomført våren 2025. I tillegg ble det testet en SMS-undersøkelse for å nå yngre aldersgrupper, en aldersgruppe som har lav «respons rate» (deltakelsesprosent) på telefonunderundersøkelser. I forsknings-prosjektet har vi også vurdert bruk av registerdata til vekting, i rekruttering til undersøkelsen, til å redusere antall spørsmål i undersøkelsen, og for å gjøre undersøkelsen mer sammenlignbar med nasjonale levekårsundersøkelser. Det langsiktige målet er et Synsbarometer med høy metodisk kvalitet, som kombinerer innsamlet informasjon fra synshemmede med relevante registerdata, samtidig som det tar hensyn til utfordringer knyttet til representativitet og skjevheter i utvalget. Vår anbefaling for Synsbarometeret 2027 er å søke SSB om å få bruke registerdata. Formålet er å bruke registerdata til uttrekket til undersøkelsen, og til analyser og postvekting etter at undersøkelsen er gjennomført.

The Norwegian Computing Center has been working on automating the inspection and monitoring of railway infrastructure for the last 5 years in collaboration with Bane NOR. Together, NR and Bane NOR are developing a mobile and cost-efficient camera system that can be mounted on the front of maintenance trains to record video of the infrastructure. We have been using deep learning–based computer vision on this data to detect faults, identify anomalies, and perform positioning for change detection. In this presentation, Anders Waldeland talks about this work and how we have been using various foundation models and self-supervised learning to advance automatic inspection.

The publication of court judgments is essential to upholding rule of law and democratic norms as well as facilitating legal research, and new legal technologies. However, many European states struggled to transition to online publication at scale. In this article we address three questions: what are the obligations of states to publish judgments; which states are making progress; and what are the challenges and solutions in ensuring greater publicity? We examine the overarching duties in the ECHR and EU law and the relevant legal requirements and practice in 12 national jurisdictions and two regional courts. Our findings show tremendous variation in duties and practice, and identify barriers to progress (legal, organisational, and budgetary) but also promising innovative solutions in certain jurisdictions. Ultimately, while this publication diversity provides a form of experimental governance, it would be timely to move towards common standards and approaches.

Federated decentralized IoT systems are reshaping how data is exchanged and processed across domains such as smart cities and healthcare, yet ensuring trust in such dynamic, distributed environments remains a significant challenge. This paper introduces FedTrust, a layered framework that integrates decentralized communication, federated data services, and a self-adaptive trust-risk model to assess the trustworthiness of IoT devices in real time. By extending an established analytical trust model, we refine existing constructs and introduce new dimensions such as context awareness and behavioral monitoring to account for the operational variability of edge devices. Our proposed model computes risk and trust scores using weighted metrics, driving automated decisions and mitigation actions via a closed feedback loop. FedTrust provides a scalable and resilient approach for securing federated IoT ecosystems through continuous, data-driven trust calibration.

Advancements in Deep Neural Network (DNN) models and hardware accelerators have made edge intelligence a practical alternative to cloud-based intelligence. However, application-specific requirements, such as accuracy, latency, security, and privacy, as well as workload fluctuations, necessitate dynamic allocation of edge and cloud resources. To facilitate such dynamic allocation, we propose an adaptive model selection and switching framework that leverages operational performance scores. We evaluate the approach using various object classification models, demonstrating its ability to balance accuracy and inference time while ensuring scalability and efficient resource utilization.