37115 research outputs found
Sort by
Sparse Neural Network Interpretability: A Comparative Analysis of Au toencoders and Transformers
Get PDFAs neural networks grow increasingly complex and powerful, understanding their internal representations becomes critical for ensuring safe and reliable AI systems. This thesis addresses a fundamental challenge in mechanistic interpretability: how architectural choices in sparse representation learning shape our ability to understand neural network internals. We present the first comprehensive theoretical and empirical comparison of Sparse Autoencoders (SAEs) and Sparse Transformers (STs), two competing approaches for decomposing neural network activations into interpretable features.
We investigate a unified geometric framework revealing that these architectures, despite solving the same sparse coding problem, operate in fundamentally different spaces. SAEs use ReLU activations to create unbounded sparse features in positive orthants, while STs employ softmax attention to confine features to the probability simplex. These constraints lead to complementary capabilities: SAEs excel at Euclidean separation through magnitude differences, while STs optimize for angular discrimination through competitive dynamics.
Our empirical validation across visual datasets (MNIST and Fashion-MNIST) confirms dramatic performance differences predicted by theory. SAEs achieve 100--1000× superior Euclidean separation between class representations, while STs demonstrate 10× better angular discrimination. Feature-level analysis reveals that SAEs learn shared, polysemantic features that capture cross-class patterns, whereas STs develop highly specialized, monosemantic features through winner-take-all competition. When applied to GPT-Neo 1.3B language model representations, these performance gaps narrow substantially (to 3--30×), suggesting that architectural biases, while persistent, are moderated by the complexity of real-world applications.
This work establishes that the choice of interpretability architecture is not merely an implementation detail but fundamentally determines what aspects of neural computation we can observe and understand. SAEs provide superior tools for identifying distinct concepts requiring clear on/off behavior, while STs excel at revealing relational structures and relative feature importance. These complementary strengths suggest that future interpretability research should focus not on choosing between architectures but on understanding how to leverage their respective advantages for comprehensive mechanistic understanding
Drømmen om samisk i norsklæreverkene
Get PDFI denne oppgaven presenterer jeg skisser for hvordan læremidlene i norskfaget kunne vært utforma for å speile læreplanverkets intensjoner om å inkludere samer i et mangfoldig «vi». Arbeidet gis form gjennom å bygge forståelse i tre deler. Hver del tar sikte på å svare på et forskningsspørsmål:
Det første spørsmålet: Hva vi vet om samiske emner i læremidler for norskfaget svarer jeg på gjennom å oppsummere hva den nyeste forskninga på samiske emner i læremidler har funnet. Dette kan sammenfattes med at det er mulig å se en gradvis nyansering i hvordan samiske emner behandles i læremidlene i norsk skole, men samiske emner har fremdeles en marginal plass i læremidlene, det preges av en norsk majoritetsforståelse av «det samiske» og det er fremdeles stort rom for forbedring.
Det andre spørsmålet: hvordan vi kan forstå utvalget og organiseringa av samisk fagstoff i læremidler i norskfaget i et språkideologisk og indigeniserende rammeverk svarer jeg på gjennom å gjøre en læremiddelanalyse av norsklæreverket Salto 5-7 fra Gyldendal forlag. Gjennom å trekke på kritisk diskursanalyse som forskningsmetode og se funnene i lys av samisk kunnskapsforståelse, Irvine og Gals (2000) språkideologiske rammeverk og et rammeverk for indigenisering av undervisning basert urfolkspedagogisk forståelse av mainstreaming/whitestreaming og andregjøring og Torjer A. Olsens (2017) kategorier fravær, inkludering og indigenisering i mainstreamundervisning får jeg en forståelse for hvordan den eksisterende strukturen i læreverket kommer i konflikt med målet om å integrere samiske perspektiver på en meningsfylt og relevant måte. Mine funn samsvarer med funnene fra tidligere forskning i at samiske emner fremdeles har en marginal plass, at det preges av en norsk masjoritetsforståelse og at det er stort rom for forbedring.
Det tredje og siste spørsmålet: hvordan læremidlene kunne vært uforma og formulert for å sikre at samiske perspektiver oppfattes som meningsfylt og relevant for samiske og ikke-samiske elever, svarer jeg på ved å gjøre en aksjon der jeg bearbeider og forfatter læremiddeltekster basert på erfaringene fra de to første delene.
Jeg konkluderer med at det er mulig å utforme læremidler som tar utgangspunkt i samiske perspektiver, som integrerer samiske emner på måter som oppfattes som meningsfylt og relevant for både samiske og ikke-samiske elever, men det krever store endringer i læremidlenes struktur og tekstpraksis
Design of an Emergency Recovery System (ERS) for High Altitude Drone Systems
No full textFull text not availableA new type of emergency parachute system is developed using engineering, combining off-the-shelf
components with parts produced with additive manufacturing. A new ultralight method for retaining
the the parachute and lid using as thin line is developed and tested, An ultralight redundant trigger
mechanism with health monitoring to ensure reliable triggering is also developed and tested. Parts and
design are evaluated through calculations, aerodynamic simulation and modeling as well as physical
testing. The prototype weighs 438 g and occupies a volume of about 1.1 Liters. Testing found that it
was possible to surpass existing solutions in terms of mass and if this design was adapted to a lighter
parachute, which is commercially available the total mass of ERS would be less than 300 g and a
volume of less than 0.6 Liters
Fådelte skolers møte med læreplanen i matematikk
Get PDFI vårt masterprosjekt har vi undersøkt hvilke utfordringer og muligheter matematikklærere ved fådelte skoler ser i møte med Læreplanverket for kunnskapsløftet 2020 (LK20). Bakgrunnen for vårt masterprosjekt er at innføringen av LK20 førte til en strukturendring i matematikkfaget. Endringen fra K06 til LK20 gikk blant annet ut på at det ble innført kompetansemål for hvert trinn i matematikk, samt innføring av nye læreplanelementer. For oss var det interessant å utforske hvordan matematikklærere ved fådelte skoler, som kjennetegnes ved aldersblanding, har opplevd implementeringen og praktiseringen av LK20. Teorigrunnlaget i vår masteroppgave tar utgangspunkt i læreplanteori, undervisningskunnskap i matematikk og fådelte skoler og aldersblanding.
For muligheten til å utforske erfaringer og opplevelser i dybden valgte vi et kvalitativt forskningsdesign. Innsamlingen av lærernes erfaringer og opplevelser ble gjort gjennom semistrukturerte intervjuer. For å helhetlig belyse tematikken i vårt masterprosjekt er en representant fra Utdanningsdirektoratet (Udir) del av vårt utvalg, i tillegg til fem matematikklærere fra én kommune. I analysering av matematikklærer-intervjuene benyttet vi stegvis deduktiv induktiv-metode (SDI) som utgangspunkt, med et behov for å avslutningsvis fullføre analysen selvstendig. Formålet med intervjuet av representanten fra Udir var å gi et nyansert bilde av masteroppgavens tematikk. Bearbeidelsen av dette intervjuet ble gjort med utgangspunkt i de allerede analyserte intervjuene fra matematikklærerne.
Vi fant at lærerne opplever flere utfordringer enn muligheter i sitt møte med LK20. Utfordringene sentreres hovedsakelig rundt at læreplanen nå har kompetansemål for hvert trinn. Denne strukturen fører til at lærerne selvstendig må finne tilnærminger for at elevene får den undervisningen de har krav på. Eksempelvis undervise elevene i flere kompetansemål enn tiltenkt og lage rullerende toårsplaner. Hvordan kjerneelementene skal praktiseres blir oppfattet ulikt av lærerne og faglitteraturen på en side, og Udir på den andre, noe som vanskeliggjør implementeringen for lærerne. Det kommer frem gjennom utfordringene og ønskene fra lærerne at fådelte skoler i liten grad er blitt tatt hensyn til i læreplanarbeidet, noe Udir selv innrømmer. Våre intervjuer fremmer hvordan et økt handlingsrom i møte med kompetansemålene i LK20 vil lette læreplanarbeidet og implementering i en fådelt kontekst. En forutsetning for at endring kan skje er at utfordringene fådelte skoler opplever løftes slik at Udir blir bevisst utfordringenes eksistens. Vårt masterprosjekt er dermed et bidrag til forbedring av læreplanen i matematikk for fådelte skoler
Artificial Intelligence in National Security Intelligence
No full textFull text not availableThis thesis examines how artificial intelligence (AI) is being integrated into national intelligence work, focusing on operational practices, institutional dynamics, and ethical implications. It addresses three main research questions: How AI is being implemented in institutions such as the Central Intelligence Agency (CIA), how Ukrainian intelligence agencies apply AI in the context of full-scale war, and how private AI actors like Palantir influence public intelligence infrastructure. The thesis uses qualitative document analysis based on policy papers, public interviews, and academic sources to explore these themes. The findings show that AI integration depends not only on technical capacity, but also on how institutions interpret and manage its application. The CIA illustrates a model of selective adoption, where human judgment remains central, and AI tools are introduced under strict legal and procedural controls. In Ukraine, AI technologies are applied across a broad range of intelligence and operational tasks, including geospatial analysis, open-source intelligence (OSINT), targeting, and information operations, demonstrating a more experimental and necessity-driven use. The Palantir case highlights how corporate platforms play a growing role in national security, raising questions about public oversight, data governance, and institutional dependency. The thesis suggests that AI in intelligence can be understood as part of a socio-technical system, where institutional norms, human roles, and technical design are interdependent. While AI offers new capacities, its influence depends on how it is embedded in organizational settings and subject to political and legal constraints. The study contributes conceptually and empirically to understanding how intelligence agencies and private firms navigate the changing landscape of digital security governance
Preserving Privacy in Interactions with Large Language Models
Get PDFIn this thesis we investigate the preservation of privacy in user interactions with Large Language Models (LLMs), focusing on transforming user queries to enhance privacy while maintaining the usability of answers. The research is contextualized within the FysBot mobile health application, which aims to motivate physical activity.
The core problem addressed is the potential leakage of sensitive user information through prompts sent to LLM-based chatbots, stemming from risks like data memorization, re-identification, and logging. This thesis proposes a privacy-preserving system designed to mitigate these risks by modifying queries before they reach the external LLM.
The developed system employs several techniques: numerical data (e.g., steps, geolocation, heart rate, time) is perturbed using randomized noise through methods like General Additive Data Perturbation (GADP) and Multiplicative Data Perturbation (MDP), tailored to the specific data type to maintain utility. Sensitive textual information is identified and substituted with semantic labels chosen via cosine similarity on text embeddings. The system was implemented in Python, utilizing models like ChatGPT 3.5 and text-embedding-3-small.
Evaluation of the system involved performance benchmarking and a user survey. Benchmarking revealed a significant overhead, with an approximate 2.3-fold increase in data sent, a 3.7-fold increase in data received, and a 3-fold increase in execution time when the privacy-preserving system was used. The user survey, conducted with participants from health research and the general public, indicated that while a vast majority preferred answers generated from original, sensitive prompts. 50% of participants were willing to accept a reduction in the usability of answers in exchange for enhanced privacy. Hesitancy was often linked to the criticality of the sensitive information (diagnoses), where accuracy was deemed paramount.
This thesis concludes that it is feasible to develop a system that enhances end-user privacy in LLM interactions with a manageable loss in usability. However, the introduced overhead suggests Backend implementation is more viable for mobile applications. Future work could focus on handling real-time sensitive data detection or optimizing the system’s performance
Multidose i e-resept - Optimalisering av praksis basert på tilbakemeldinger fra helsepersonell
Get PDFAbstrakt
Bakgrunn: Multidose i e-resept (eMD) er en digital helseløsning utviklet for å forbedre legemiddelhåndtering, pasientsikkerhet og samhandling i helsetjenesten. Til tross for dette viser praksis og forskning at løsningen fører med seg betydelige utfordringer ved implementering og bruk. Om lag 85 000 pasienter skal overføres fra papirmultidose til elektronisk multidose i løpet av de kommende årene, noe som forsterker behovet for en trygg og effektiv løsning og gjør problemstillingen i denne studien særlig aktuell.
Problemstilling: Studien undersøker hvilke utfordringer leger og farmasøyter møter i arbeid med eMD, samt hvilke løsninger de selv foreslår for å forbedre praksis.
Metode: Det ble benyttet et mixed methods-design med parallell innsamling av kvantitative og kvalitative data. Et strukturert litteratursøk ble gjennomført som grunnlag for utvikling av et spørreskjema, som ble pilotert og sendt ut til 54 apotek og 190 legekontor i Norge. Innsamlede data ble analysert ved hjelp av tematisk analyse.
Resultater: Respondentene opplever utfordringer med uklar ansvarsfordeling, mangelfull opplæring, svak teknisk støtte og utilstrekkelig samhandling. Systemet beskrives som tidkrevende og ressurskrevende, med risiko for dobbeltforskrivning og feilutlevering. Manglende oversikt over legemiddellister og svak systemintegrasjon øker faren for feilmedisinering. Det etterlyses tydeligere roller, bedre støttefunksjoner, mer praktisk opplæring og teknologiske forbedringer.
Konklusjon: Studien viser at eMD har stort potensial, men at videreutvikling krever innsats på tvers av teknologiske, organisatoriske og faglige nivåer. Helsepersonellets innsikt er verdifull og bør aktivt benyttes i den videre utviklingen av løsningen før bredere utrulling.Abstract
Background: Electronic multidose (eMD) is a digital healthcare solution designed to improve medication management, patient safety, and collaboration in the health services. Despite its aims, practice and research show that the system presents considerable challenges during implementation and use. With approximately 85.000 patients transitioning from paper-based to electronic multidose, there is an urgent need for a safe and efficient solution, making this study highly relevant.
Research question: This study examines the challenges faced by physicians and pharmacists when using eMD and explores their suggested improvements.
Method: A mixed methods design was applied, combining quantitative and qualitative data collection. A structured literature review informed the development of a questionnaire, which was piloted and distributed to 54 pharmacies and 190 general practices in Norway. The data were analysed thematically.
Results: Respondents report unclear responsibility, insufficient training, limited technical support, and weak collaboration. The system is seen as time-consuming and resource-intensive, with risks such as duplicate prescribing and medication dispensing errors. Poor overview of medication lists and weak integration between systems increase the likelihood of medication errors. Respondents call for clearer role definitions, stronger support systems, more hands-on training, and improved technological infrastructure.
Conclusion: eMD holds significant potential, but its development requires coordinated efforts across technological, organizational, and professional domains. The experiences of healthcare professionals are essential and should be actively included in the ongoing development process before wider implementation
Southern Ocean Carbon Export Revealed by Backscatter and Oxygen Measurements From BGC-Argo Floats
Get PDFThe Southern Ocean (south of 30°S) contributes significantly to global ocean carbon uptake through the solubility, physical and biological pumps. Many studies have estimated carbon export to the deep ocean, but very few have attempted a basin-scale perspective, or accounted for the sea-ice zone (SIZ). In this study, we use an extensive array of BGC-Argo floats to improve previous estimates of carbon export across basins and frontal zones, specifically including the SIZ. Using a new method involving changes in particulate organic carbon and dissolved oxygen along the mesopelagic layer, we find that the total Southern Ocean carbon export from 2014 to 2022 is 2.69 ± 1.23 PgC y−1. The polar Antarctic zone contributes the most (41%) with 1.09 ± 0.46 PgC y−1. Conversely, the SIZ contributes the least (8%) with 0.21 ± 0.09 PgC y−1 and displays a strong shallow respiration in the upper 200 m. However, the SIZ contribution can increase up to 14% depending on the depth range investigated. We also consider vertical turbulent fluxes, which can be neglected at depth but are important near the surface. Our work provides a complementary approach to previous studies and is relevant for work that focuses on evaluating the biogeochemical impacts of changes in Antarctic sea-ice extent. Refining estimates of carbon export and understanding its drivers ultimately impacts our comprehension of climate variability at the global ocean scale
The Role of Ballasting, Seawater Viscosity and Oxygen-Dependent Remineralization for Export and Transfer Efficiencies in the Global Ocean
Get PDFThe particulate organic carbon (POC) flux from the euphotic zone to the deep ocean is central to the biological carbon pump. It is typically evaluated using “export efficiency” and “transfer efficiency,” which reflect POC formation and sinking and carbon sequestration efficiency in the ocean's interior, respectively. Since observations of these metrics are limited, biogeochemical models can elucidate the controls of large-scale patterns. This study uses the global ocean-biogeochemical model FESOM-REcoM, with a new sinking routine that accounts for ballast minerals, seawater viscosity, and oxygen-dependent remineralization in POC sinking and remineralization, to identify the drivers of global export and transfer efficiency. We find that export efficiency is highest at high latitudes, where diatoms, mesozooplankton, and macrozooplankton dominate the plankton community, but that high export efficiency does not always imply high transfer efficiency. Omitting ballast minerals decreases export efficiency by 20% in the Southern Ocean, yet the globally integrated POC flux out of the euphotic zone (5.4–5.6 Pg C yr-1
) and the global average export efficiency (14.7%–15.4%) are relatively insensitive to seawater viscosity, mineral ballasting, or oxygen-dependent remineralization. In contrast, global transfer efficiency is more sensitive to these processes and varies between 21% and 25% in the simulations, with the largest reduction by 23% observed when omitting ballasting in subtropical, low-productivity regions. Our findings suggest that assumptions about ballasting and background sinking speed could explain previous discrepancies in the literature regarding the highest transfer efficiencies in low or high latitudes. Notably, while plankton community structure determines export efficiency regimes, zooplankton fecal pellets drive high transfer efficiencies in regions with high export efficiency, like the Southern Ocean
A national outbreak of Serratia marcescens complex: investigation reveals genomic population structure but no source, Norway, June 2021 to February 2023
Get PDFWe report a national outbreak of Serratia marcescens complex type 755 (ct755) in Norway, with 74
cases identified between June 2021 and February
2023. Careful reviews of patient journals and
interviews were performed, involving 33 hospitals
throughout Norway. All available clinical isolates
of S. marcescens collected between January 2021
and February 2023 (n=455, including cases) from all
involved hospitals were whole genome sequenced.
Cases displayed a pattern of opportunistic infections,
as usually observed with S. marcescens. No epidemiological links, common exposures or common risk
factors were identified. The investigation pointed to
an outbreak source present in the community. We suspect a nationally distributed product, possibly a food
product, as the source. Phylogenetic analysis revealed
a highly diverse bacterial population containing multiple distinct clusters. The outbreak cluster ct755 stands
out as the largest and least diverse clone of a continuum, however a second cluster (ct281) also triggered
a separate outbreak investigation. This report highlights challenges in the investigation of outbreaks
caused by opportunistic pathogens and suggests that
the presence of identical strains of S. marcescens in
clinical samples is more common than previously
recognised