Velg hva du ønkser å vise prosjekt for.
This master thesis is part of the newly established AI Centre for the Empowerment of Human Learning (AI LEARN) — one of six national AI centers in Norway.
[ Vis hele beskrivelsen ]
This thesis will be co-supervised by KS. KS is an interest organization, employer organization and development partner for all municipalities and county authorities.
Geological field trips are an essential component of geoscience education and research, allowing students and researchers to observe geological formations, collect samples, and document field conditions. Traditionally, geological data collected during fieldworks, such as photographs, notes, GPS coordinates, and measurements, are analyzed later using laboratory or desktop tools. This workflow can delay feedback and limit opportunities for immediate learning and interpretation in the field.
Recent advances in Edge AI enable data analysis directly on mobile devices such as smartphones, tablets, or portable field computers. By processing geological images and observations locally, students and researchers can receive immediate assistance in identifying rock types, geological structures, or stratigraphic features. At the same time, more complex analysis tasks can be handled by cloud systems once connectivity is available.
The concept of a cloud–edge computing continuum allows geological data processing to be distributed between edge devices in the field and cloud infrastructure. This approach can improve responsiveness during fieldwork while still enabling advanced analysis and data aggregation.
This thesis aims to investigate how Edge AI combined with a cloud computing continuum can support geological fieldwork and improve learning and data analysis during geology field trips.
The results may help support the development of intelligent digital tools for geoscience education and field research.
[ Skjul beskrivelse ]
Digital learning environments are increasingly used in schools to support language learning through interactive exercises, speech analysis, and adaptive learning platforms. These systems generate large amounts of student interaction data that can be used to personalize learning experiences and provide timely feedback. However, sending all student data directly to centralized cloud services raises concerns related to latency, privacy, and network dependency.
Edge AI offers a promising alternative by enabling data processing directly on local devices such as tablets, laptops, or classroom servers. By performing initial analysis at the edge, sensitive student data can remain closer to its source while still allowing intelligent feedback and adaptive learning. At the same time, more computationally intensive tasks can be handled in the cloud. The concept of a cloud–edge computing continuum allows tasks to be distributed dynamically between local and remote computing resources.
This thesis aims to investigate how Edge AI can be integrated with a cloud computing continuum to support efficient and privacy-aware analysis of student data in language learning environments.
The results may help support the development of scalable, privacy-aware educational technologies that improve personalized language learning experiences.
Training for high-risk situations such as emergency response, disaster management, military operations, and industrial safety often requires realistic simulations. Traditional training methods, including physical exercises and manual scenario evaluation, can be costly, difficult to scale, and limited in their ability to provide detailed performance analysis.
Serious games and simulation-based training environments have emerged as promising tools for preparing individuals for such scenarios. However, evaluating trainee behavior within these environments typically relies on predefined rules or manual observation, which can limit the depth and accuracy of feedback.
Advances in computer vision and vision-based artificial intelligence (Vision AI) offer new opportunities to analyze gameplay and simulation recordings automatically. By interpreting visual information from game screens or simulation environments, Vision AI systems can detect actions, events, and environmental states in real time or from recorded sessions. This enables automated assessment of trainee performance and deeper analysis of decision-making under high-risk conditions.
This thesis aims to explore how Vision AI can be used to extract meaningful data from training simulations or games to support performance analysis and feedback.
The results may support the development of more intelligent and scalable training systems for safety-critical environments.
This master's thesis is part of the newly established AI Centre for the Empowerment of Human Learning (AI LEARN) — one of six national AI centers in Norway.
We are looking for motivated Computer Science master’s students to join an innovative thesis project exploring how artificial intelligence can transform the development process of game-based learning (GBL) applications and serious games.
The goal of this project is to investigate how AI can make the development of GBL applications faster, more efficient, and less resource-intensive. Students will explore how AI can support or automate parts of the development pipeline, such as:
The project can take multiple directions depending on student interests, including:
The project will combine technical development with scientific investigation. In addition to building and/or designing solutions, students will:
This is a research-driven, forward-looking project ideal for students interested in AI, software engineering, game development, or educational technology. The outcome may include prototypes, tools, frameworks, or design guidelines that contribute to more accessible and scalable development of learning games.
This project requires two students to work in a group, but multiple groups can work on variants of this project.
Students will gain valuable experience in AI-driven development, tool design, and applied research, preparing them for careers in both industry and academia.
Artificial intelligence models such as large language models (LLMs), computer vision systems, and recommendation models are increasingly integrated into modern software applications. Many organizations now rely on multiple AI services simultaneously, including both internal models and external APIs. However, direct interaction between applications and AI models often leads to challenges related to scalability, monitoring, cost management, security, and governance.
One emerging architectural solution is the AI Gateway. Similar to an API gateway in microservice architectures, an AI gateway acts as an intermediary layer that manages communication between applications and AI models. Such a gateway can provide functionality including request routing, access control, monitoring, prompt processing, and model selection. Despite its growing importance in industry AI platforms, the concept of AI gateways has received limited systematic study in academic research.
This thesis aims to investigate the design and implementation of an AI gateway that facilitates efficient, secure, and scalable interaction with AI models.
We are looking for motivated Computer Science master’s students to join an exciting thesis project at the intersection of artificial intelligence, game development, and learning science.
The goal of this project is to design and develop an innovative game-based learning application enhanced by AI. Students will explore how AI can be leveraged to:
The project will combine technical development with scientific investigation. In addition to building the application, students will:
This is a hands-on, research-driven project ideal for students interested in AI, game development, HCI, or educational technology. The outcome will be both a functional prototype and experimental results contributing to ongoing research in intelligent learning systems.
Students will gain experience in AI integration, game design, user testing, and academic research, making this project highly relevant for both industry and research careers.
This thesis is co-supervised by Levato and Sjøkrigsskolen.
Modern digital games produce large-scale game telemetry data, capturing detailed logs of player actions, sessions, and progression across time. These datasets are inherently hierarchical, with repeated observations nested within players, sessions, and game levels. However, many traditional analytical approaches treat such data as flat, ignoring these nested structures and thereby violating key statistical assumptions, particularly independence of observations. This can lead to biased estimates and misleading conclusions, especially when attempting to understand player behavior, engagement, or progression patterns.
Multilevel modeling (also known as hierarchical linear modeling) provides a principled framework for analyzing such data. By explicitly accounting for nested relationships, these models can separate within-player variation (e.g., how a player behaves across sessions) from between-player variation (e.g., differences in skill or playstyle). Furthermore, they enable the analysis of how contextual factors, such as level design, session conditions, or gameplay features, influence player behavior over time.
This thesis aims to explore the application of multilevel statistical models to game log data, providing more accurate and interpretable insights into player behavior and game design dynamics.
Research Objectives
Potential Impact
We are looking for two motivated students to join an industry-linked thesis project in collaboration with Kahoot!, focusing on the role of artificial intelligence in modern learning platforms.
The goal of this project is to investigate how AI is currently used within the Kahoot! platform and to identify opportunities for improvement and innovation. Students will analyze existing AI-driven features, such as the AI content generator and related functionalities, and explore how these can be enhanced to improve learning quality, usability, and engagement.
The project will involve:
The project combines technical insight with user- and product-oriented research. Students will:
This project is conducted in collaboration with Kahoot!, giving students a unique opportunity to work with a real-world product used globally. The project will be supervised by Alf Inge Wang (inventor/co-founder of Kahoot!), and students will have a dedicated contact at Kahoot! for product-specific guidance and insights.
This is an excellent opportunity for students interested in AI, HCI, software engineering, game-based learning and educational technology, particularly those who want experience working closely with industry.
Students will gain hands-on experience in AI evaluation, product analysis, and applied research in an industry context, making this project highly valuable for future careers in both academia and the tech industry.
Game analytics traditionally relies on internal telemetry data collected directly from game engines. While highly informative, such data is often inaccessible, especially in cases involving proprietary systems, legacy games, or third-party analysis. At the same time, vast amounts of gameplay recordings are publicly available through platforms such as streaming services, esports broadcasts, and user recordings. These videos contain rich information about player actions, environment states, and game dynamics, but remain largely underutilized due to the lack of structured data extraction methods.
Recent advances in computer vision and deep learning provide an opportunity to bridge this gap. By treating gameplay footage as a sequence of visual observations, it is possible to automatically detect and interpret key elements such as player movements, UI changes, events, and interactions.
This thesis aims to develop a vision-based AI framework that transforms raw gameplay video into structured, analyzable data, enabling new forms of game analytics without requiring access to internal game logs.
Expected Contributions
This umbrella project envisions the design and development of a comprehensive AI-driven ecosystem for language learning, centered on enhancing human learning through intelligent, adaptive, and privacy-conscious technologies. The project aligns with the goals of the AI LEARN centre, focusing on how AI can meaningfully support and empower learners through interaction, personalization, and multimodal experiences.
At its core, the project aims to integrate multiple AI paradigms, such as, conversational AI, speech processing, computer vision, multimodal learning, and learning analytics, into a unified framework that supports learners across different dimensions of language acquisition: speaking, listening, vocabulary, comprehension, and progression over time.
Here are some short descriptions of example projects.
The umbrella project addresses a central question: How can AI systems be designed to provide effective, engaging, and ethical language learning experiences that adapt to individual learners while preserving privacy?
To answer this, the project combines:
Expected Impact
This umbrella project focuses on developing a next-generation AI ecosystem for programming education at the university level, aimed at supporting deeper learning, critical thinking, and effective collaboration. Aligned with the vision of the AI LEARN centre, the project explores how AI can move beyond automation to become a pedagogically aligned partner in learning, enhancing rather than replacing student effort.
The project brings together advances in code understanding, generative AI, debugging support, collaboration analysis, and multimodal learning analytics to create intelligent systems that support students throughout the programming lifecycle, from writing and debugging code to collaborating and reflecting on their learning processes.
The umbrella project addresses a central question: How can AI systems be designed to support programming education in higher education in ways that enhance deep learning, maintain student agency, and provide meaningful insight into both code and learning processes?
To address this, the project combines:
This umbrella project aims to design and evaluate a comprehensive AI-powered ecosystem for programming education, with a particular focus on K–12 learners and beginners. Grounded in the vision of the AI LEARN centre, the project explores how human-centered AI can make programming more accessible, engaging, and pedagogically effective through personalization, guidance, and interaction.
The project integrates advances in conversational AI, program analysis, generative AI, visualization, and educational scaffolding to support learners across the full programming journey, from understanding basic concepts to solving problems, debugging code, and connecting abstract logic to real-world systems.
The umbrella project addresses a central question: How can AI systems be designed to support programming education in ways that are adaptive, explainable, and engaging; while fostering deep understanding rather than shortcut learning?
This umbrella project envisions a comprehensive AI-powered ecosystem for writing education, aimed at supporting students throughout the entire writing process—from idea generation to revision and assessment. Anchored in the vision of the AI LEARN centre, the project focuses on how AI can enhance writing as a cognitive and creative process, without replacing the learner’s own thinking and expression.
The project integrates advances in large language models, explainable feedback systems, multimodal AI, and educational design to create tools that provide guided, interactive, and pedagogically aligned support for diverse learners, including K–12 students and ESL learners.
The umbrella project addresses a central question: How can AI systems be designed to support writing as a guided, creative, and reflective process, enhancing learning while preserving student agency and originality?
To explore this, the project combines:
Open to proposing your own topic in the intersection of Human-Centered AI (HCAI) and Learning Technologies.
There is an option to collaborate with different EdTech companies and public sector partners from the AI LEARN centre. For example, LearnLab, Kahoot!, Visma Flyt, KS, and Sikt.
Supervisors: Michalis Giannakos, Giulia Cosentino, Noah Sejer Iversen
Brief Description
This master's thesis is part of a broader research collaboration between NTNU, LearnLab, and ongoing Participatory Design (PD) studies with Norwegian teachers and students. Grounded in PD, the project explores how teachers can actively and meaningfully participate in shaping AI-powered learning tools through dialogue, mutual learning, and structured, interpretable feedback for hybrid human-AI models.
Description of the work
As AI-supported learning platforms become increasingly embedded in schools, an important challenge is to ensure that teachers are not reduced to passive overseers of systems they did not help shape. Inspired by PD traditions of democracy at work, mutual learning, and the collective shaping of technology, the thesis explores how teachers can take an active role in influencing how AI-based feedback systems operate in educational settings. More specifically, the thesis aims to design, prototype, and evaluate a hybrid human-AI feedback model through which teachers can guide, negotiate, correct, and modulate AI-generated feedback to students. This includes developing algorithms and interfaces that make teacher input an integrated and meaningful part of how the AI system behaves.
Possible directions include, but are not limited to:
The exact direction, scope, and implementation details will be agreed upon together with the supervisor, based on the student’s interests, motivation, and the needs of the platform. The thesis may include the design, development, and evaluation of one or more models.
Requirements
The ideal students should have experience in software design or AI/ML and an interest in educational technologies.
Recommended technical skills: Backend/API development (Python), basic ML, frontend prototyping, and optional experience with UX or PD. Expected outcomes might include:
Brief Description:
Multi-Agent Systems (MAS) consist of multiple interacting agents, each with a specific role and function within the system. This area has recently attracted significant research interest. These systems can effectively manage large-scale and complex data through collaboration and task distribution. The field of education requires collaborative effort to achieve its goals; therefore, the development of MAS represents a promising solution for implementing an educational system capable of providing accurate and fast responses, encouraging meaningful dialogue between students, teachers, and the system itself. This study investigates whether MAS can effectively support different educational use cases. Specifically, this thesis aims to design MAS using a combination of rule-based scripts and LLM APIs and to evaluate their performance through assessments conducted by human experts.
Description of the work
The thesis will investigate how MAS can support teaching and learning. The work will focus on designing MAS and may address pedagogical, technical, or AI-related challenges. Specifically, the aim of this work is to examine the extent MAS can be useful in various educational scenarios and to understand how these systems influence people through their interaction with students/teachers.
The exact direction, scope, and implementation details will be agreed upon together with the supervisor, based on the student’s interests and motivation. The thesis may include design, development, and evaluation of MAS.
The ideal candidates will have a background in software design, solid programming skills, and an interest in hands-on development and experimentation.
Recommended technical skills:
Large language models (LLMs) have demonstrated strong capabilities in educational applications, supporting learning across multiple domains, including language and mathematics. However, their substantial number of parameters results in high computational demands, making them difficult to deploy in educational environments with limited infrastructure. Consequently, small language models (SLMs) have recently gained attention as a more efficient and accessible alternative. This study investigates whether SLMs can effectively support educational use cases. Specifically, the thesis aims to design and conceptualize multiple SLM pipelines using diverse prompt templates and fine-tuning strategies, and to evaluate their performance through assessments conducted by human experts.
The thesis will investigate how SLMs can be developed to support teaching and learning. The work should be connected to the design and enhancement of SLMs and may address pedagogical, technical, or AI-related challenges. In particular, the aim of this work is to examine whether SLMs can achieve useful performance on different reasoning tasks while operating under limited computational resources. By fine-tuning open-source models on domain-specific datasets, this work seeks to lower the barrier for applying AI in relevant applications and to encourage further exploration of SLMs in resource-constrained educational environments (e.g., run models locally on mobile devices).
IntroductionThis thesis aims to investigate how interactive mobile technology (mobile application) enhanced with LLMs and conversational agents (e.g., open source examples CodeLLama, ParlAI, ChatterBot), as well as (Multimodal) LLMs can support learning.
Thesis DescriptionThe student(s) need to review the literature and familiarize themselves with relevant apps and AI technology. The focus is to integrate (M)LLM and conversational agent mechanics (e.g., allow users to interact) in an engaging way. Based on the best practices from the literature, the candidates will develop an app (e.g., following co-design or participatory design) and then do a user study either in school settings or an informal learning arena to empirically test the proposed system. Finally, the candidate will analyze the collected data and write up their thesis.
RequirementsThe ideal candidate will have a background in user experience, interface design, and use/integration of LLMs. Solid front-end programming skills (JavaScript and CSS) and an interest in hands-on development and experimentation.
Programming skills: MySQL, JavaScript, CSS.
Examples of previous theses:
https://hdl.handle.net/11250/3160949
https://hdl.handle.net/11250/3159709
Modern digital games generate vast amounts of fine-grained behavioral data through event logs, including player actions, progression, session duration, and interactions with game systems. While this data is widely used in analytics, most current approaches rely on correlational methods, which limit the ability to draw conclusions about why certain design elements influence player behavior. Understanding causality is critical for game developers, particularly in game design iteration and live-operations (live-ops), where decisions about reward systems, difficulty balancing, or onboarding experiences can significantly affect player engagement, retention, and monetization.
Recent advances in causal inference and machine learning offer new opportunities to move beyond descriptive analytics toward causal explanations of player behavior. By applying techniques such as counterfactual reasoning, causal graphs, and quasi-experimental methods, it becomes possible to isolate the true effects of design interventions while accounting for confounding factors such as player skill, playtime patterns, or prior experience.
This thesis aims to explore how causal analysis of game log data can support more informed and reliable game design decisions.
Efficient Vision Transformers for Real-Time Object Segmentation and Deployment on NVIDIA AGX Orin
Real-time object detection and semantic segmentation demand models that deliver high accuracy while remaining computationally efficient.
Vision Transformers (ViTs) have revolutionised computer vision by introducing self-attention mechanisms, which enable superior feature extraction compared to convolutional neural networks (CNNs). However, their inherent computational complexity poses challenges for edge deployment, particularly on resource-constrained platforms like NVIDIA Orin.
Unlike CNN-based models such as RTMDet, which rely on hierarchical convolutional feature extraction, ViTs perform global attention-based processing, making them more computationally intensive. Therefore, their optimisation strategies must focus on reducing redundant computations in the self-attention mechanism rather than traditional CNN pruning and quantisation techniques.
This thesis investigates transformer-specific optimisation strategies to enhance the efficiency of ViT-based object detection and segmentation models for real-time applications on NVIDIA Orin. The focus is on reducing memory footprint, improving latency, and maximising energy efficiency without sacrificing accuracy.
Research Challenges
Deploying ViT-based models on the NVIDIA AGX Orin platform presents unique challenges, such as:
- High computational costs due to self-attention mechanisms, leading to increased latency.
- Memory bandwidth constraints limiting real-time inference performance.
- The need for adaptive processing strategies to optimise transformer computations for edge deployment.
While Orin integrates GPU and DLA cores, optimal performance requires strategic allocation of computations. Unlike CNN-based architectures, which can benefit from structured pruning and layer compression, ViT optimisation necessitates attention sparsification, token reduction, and efficient matrix operations. This thesis develops transformer-specific optimisation techniques that enhance real-time ViT performance on Orin while maintaining model integrity.
Objectives
- Identify computational inefficiencies in ViT-based models when deployed on NVIDIA Orin.
- Implement and benchmark optimised ViT models on Orin, assessing improvements in speed, power consumption, and accuracy.
- Investigate optimisation techniques tailored to transformer-based architectures, including token reduction, sparsity strategies, and attention map compression.
- Develop a hybrid framework that efficiently balances computational demand and accuracy for real-time applications.
Methodology
Literature Review
- Examine research on ViT-based object detection, transformer model optimisations, and NVIDIA Orin hardware capabilities.
- Compare ViT optimisation strategies with CNN-based techniques such as those applied to RTMDet.
Baseline Implementation
- Deploy an existing ViT-based model on NVIDIA Orin using TensorRT and CUDA.
- Profile its performance, measuring inference speed, memory usage, and energy consumption.
Optimisation Techniques
- Sparse Attention Mechanisms: Implement sparse self-attention to reduce computational overhead while maintaining long-range dependencies.
- Token Pruning & Merging: Reduce token count dynamically, preserving essential information while improving efficiency.
- Mixed-Precision Quantisation: Apply FP16 and INT8 quantisation to accelerate inference and decrease memory consumption.
- GPU-DLA Offloading: Strategically distribute transformer computations between Orin’s GPU and DLA cores for efficient parallel execution.
- Memory-Efficient Self-Attention: Optimise memory access patterns, leveraging kernel fusion and hierarchical token processing to reduce latency.
- Efficient Patch Processing: Introduce adaptive patch selection and merging strategies to focus computation on informative regions.
Benchmarking and Evaluation
- Assess optimised ViT models against the baseline in terms of Latency reduction, Power efficiency, Accuracy retention.
- Compare ViT-based optimisations to CNN-based methods used for RTMDet to highlight distinct architectural benefits and limitations.
Hybrid Optimisation Framework
- Design a unified approach that combines transformer-specific optimisations to maximise real-time efficiency on NVIDIA AGX Orin.
- Ensure seamless integration into Maritime Robotics’ SeaSight pipeline.
- Provide a comparative framework for deploying ViT and CNN-based models, enabling flexible system adaptation based on specific maritime surveillance needs.
1. A thorough performance analysis of ViT-based object detection and segmentation models on NVIDIA Orin, along with a comparative study of fundamental differences in optimizations between ViT and CNN-based models.
2. A suite of optimisation techniques designed explicitly for transformer-based architectures in edge deployments.
3. Optimised ViT implementation enabling real-time inference with reduced energy consumption, seamlessly integrated into the SeaSight system for real-time situational awareness with minimal computational cost.
Tools and Resources
Hardware: Maritime Robotics’ SeaSight module, NVIDIA Jetson AGX Orin / Jetson Orin NX.
Frameworks: PyTorch, ONNX, NVIDIA TensorRT.
Profiling Tools: NVIDIA Nsight Systems, TensorRT Profiler, Jetson Power Estimation Tool.
Datasets: Custom-made dataset (or COCO, Pascal VOC for benchmarking object detection and segmentation models).
Large language models are today used in a wide range of applications like AI assistants, translation software and summarization systems. The NorwAI research center trains and publishes the largest Norwegian language models in close collaboration with the National Library, Schibsted Media and NRK. So far these models have been tested on textual use cases from the media industry, though we are now fine-tuning models for other domains and other tasks.
An interesting application of language models is speech synthesis. There are already systems available for the synthesis of standard Norwegian bokmål/nynorsk. We would like to train a system that can generate spoken dialect with the help of NorwAI's language models and a speech synthesis tool from the University of Tromsø.
The project involves a number of activities:
Experts from both NorwAI and the University of Tromsø will supervise the work.
Contact person: Jon Atle Gulla (jag@ntnu.no)
Background
Over the past decade, NTNU has developed concepts and technology for autonomous – marine vessels, and in 2019, the spinoff company Zeabuz was created to commercialize NTNU’s research on maritime autonomy. Together with the Norwegian transport company Torghatten, Zeabuz has establish the world’s first commercial route with an autonomous urban passenger ferry in Stockholm in June 2023.
Zeabuz is also developing technology for automation and autonomy in the workboat segment and are currently testing this technology on a new electric workboat developed by Yinson Greentech, intended for operation in the Singapore harbour area. This harbour is the second most trafficked harbour in the world, and has several thousand workboats such as tugs, crew-transfer vessels, pilot vessels and supply-vessels that are servicing the fleet of merchant ships visiting the harbour. To operate autonomously in this area, the autonomy system needs to interact smoothly with other traffic and comply with the traffic rules that apply.
The thesis proposal is given as a collaboration between NTNU and Zeabuz. This enables the candidate to work on a problem that is relevant for to the industry today, while also exploring methods and algorithms from the cutting edge of science. Zeabuz has several simulation environments for developing and testing motion planning and situational awareness algorithms in isolation and as an integrated part of an autonomous operation. The experimental platforms of NTNU also makes it possible to conduct full-scale experiments to test and demonstrate the real-life performance of the developed algorithms.
Motivation
Simultaneous Localization and Mapping (SLAM) is a fundamental challenge in autonomous marine navigation. Unlike terrestrial environments, marine settings can lack distinct visual landmarks or have them far away, making traditional SLAM techniques less effective. In the critical phases of autonomous operation, such as docking, precise localization is critical. SLAM can serve both as an enhancement of traditional GNSS-based navigation and as a fallback solution in case of dropouts. In addition, SLAM can also be used to accurately estimate the position of the dock in case of tidal movements etc.
Scope
This project focuses on integrating data from one or more sensors such as LiDAR, automotive radar and camera to develop a robust SLAM system for use in the docking phase of autonomous maritime navigation.
The candidate will be part of shaping the scope of this project. Some possible areas of exploration are:
Real-world testing of the developed method in cooperation with NTNU/Zeabuz is expected and cooperation with students working on control/path planning is possible. The project scope includes the project and master thesis, with a shared overarching goal: developing and evaluating a SLAM system for autonomous maritime docking. While the technical focus may vary depending on the candidate’s interests, the following outline gives a general impression of the workflow.
Project thesis:
The project thesis lays the foundation for the master's thesis by exploring the problem space and establishing a baseline solution. Expected activities include:
Master thesis:
Building on the results from the project thesis, the master thesis aims to refine the SLAM solution and demonstrate it in a real-world setting. If the baseline method shows satisfactory performance, the work will continue with:
Prerequisites
This task is comprehensive and requires a dedicated candidate with high motivation and the requisite knowledge. The candidate should be self-driven and structured and motivated to work in this field. In return we offer close supervision with bi-weekly meetings and additional follow-up when needed. This project also gives the opportunity to work on a real-life use-case with the possibility testing and experiments on a vessel platform based towards the end of the master’s thesis work. The candidate should be familiar with sensor fusion and SLAM and have good programming skills in Python or C++.
This is a project targeting selected AI-related aspects of a Situation Awareness System for an autonomous ferry, and can be performed in close cooperation with an industrial partner.
The project builds on a solide of intermediate results obtained in dozens of MSc theses supervised or co-supervised since 2022 in the “Maritime Perception Cluster” formed by R.Mester, E.Brekke, and A.Stahl
"Situation Awareness" deals with the process of processing sensor data from different sensor modalities on a moving system, in the present case: an autonomous passenger ferry and building a valid dynamic representation of the environment around the regarded mobile system. Such a dynamic representation allows to navigate safely, avoid colliisions with static obstacles and other moving vehicles, and is the basis for performing complex maneuvers such as docking a ship.
The project leaves ample space for focusing on different aspects of situation awareness and AI-based processing of sensor data, depending on the interest of the student(s) and his/her/their pre-knowledge.
The work could be focusing on typical machine learning aspects, or on "classical" (in particular statistical model-based) methods.
Most people following the recent development of AI-based computer vision will know the family of detectors running under the label “You Look Only Once” (YOLO). There detectors are really powerful and fast, and are very widely used. But they come with a systematic flaw: when applied to video streams, they usually produce flickering, unstable results, and often also multiple detections on the same object.The purpose of the project proposed here is to eliminate this flaw and let a time-aware version of YOLO which systematically builds on the history of earlier detections and generates a smooth, reliable, and temporally stable sequence of detections (bounding boxes) and segmentation results (object masks).The approach taken in this approach fuses modern machine learning models and temporal statistic processes. We will use classical (statistical) detection theory and join the insights from this theory with the learning-based approaches used in modern AI-based detectors.
This project is well suited for a student who is both interested in latest approaches from deep learning as well as proficient in dealing with the math of stochastic processes and detection theory (a field that is e.g. fundamental for radar, astronomy, and medical imaging).
Pre-Project / MSc Project Descriptions for period H2025-V2026Proposed by the Salmon Health Tracking Research ClusterDr. Christian Schellewald, Prof. Annette Stahl, Prof. Rudolf Mester, Espen HøgstedtStatus: April 2025BackgroundNorwegian salmon fish farming has established over the last few decades the world's most efficient fish production systems, and is today characterized by innovative and technology-driven production methods. Research has been and is still central for crucial advances and development of these methods.In particular as the aquaculture industry is transitioning its production methods from manual operations and experience-based reasoning towards automated and objective measurable methods using artificial intelligence and advanced mathematical models.Using cameras as intelligent sensors is crucial for moving towards more autonomous systems in different stages of aquaculture production systems. In the proposed Master-thesis projects we therefore wish to develop and exploit state-of-the-art Artificial Intelligence (AI) methods including machine learning approaches like deep-learning and other advanced methods in Computer Vision for Aquaculture applications in a new and innovative way. The students will work closely with the Aquaculture Technology collaboration team established between NTNU and SINTEF. The work is performed in the frame of the project cAIge funded by Norsk Forskningsrådet (NFR).
Topic 1: AI-based Learning for Enhanced Unsupervised Fish Welfare Indicator Detection
1.1 Background
Aquaculture, a rapidly expanding industry, necessitates an accurate monitoring of fish welfare, traditionally performed manually through observation and camera surveillance. As societal concern for animal welfare increases, the demand for sophisticated, automated monitoring solutions grows. Current methods, depending on labor-intensive manual annotations, fall short of offering the flexibility and accuracy required for large-scale operations. This project proposes the utilization of advanced AI and machine learning techniques, specifically unsupervised learning, to autonomously detect signs of deviation indicating distress or ill health in salmon, directly from video feeds. By automating the detection of welfare indicators, this approach aims to revolutionize welfare assessments in aquaculture, aligning with contemporary ethical standards.
Objective:
Develop an enhanced unsupervised and AI-based learning framework for automatic detection of welfare indicators in salmon, focusing on high-level features and semantic units.
Goals:
Anticipated Tasks:
====================================================
Topic 2: AI-based Self-learning of Semantic Segmentation of Salmon Exploiting Stereo Vision
2.1 Background/Introduction
The integration of stereo cameras in aquaculture presents a novel opportunity for enhanced monitoring, enabling the capture of 3D metric data absent in traditional video streams. This additional dimension of data may allow to self-learn an accurate semantic segmentation process of fish in videostreams, crucial for detailed welfare and health monitoring of salmon. The project exploits the depth information from stereo vision to refine self-learning algorithms for semantic segmentation, facilitating precise identification and analysis of salmon body parts in a 3D context. This advancement has the potential to provide aquaculture practitioners with unparalleled insights into fish behaviour, health and welfare, leading to improved care and management practices.
Create a self-learning framework for accurate semantic segmentation of salmon using stereo vision, enhancing understanding of fish behaviour and subsequently welfare.
===================================
Topic 3: Automated AI-based Behavioral Analysis for Salmon Welfare Assessment
3.1 Background/Introduction
The analysis of fish behavior, both at individual and collective levels, is increasingly recognized as an important component of welfare assessment in aquaculture. Behavioral anomalies may serve as early indicators of stress, disease, or discomfort. This project focuses on the development of an automated, AI-powered framework for the (real-time?) analysis of behavioral patterns, such as schooling behavior, abrupt motion changes, and feeding routines. By systematically analyzing video data to detect behavioral deviations, the proposed system aims to offer a proactive approach to identifying and addressing potential welfare issues, contributing to the ethical and sustainable management of aquaculture operations.
3.2 Objective:
Develop an automated AI-based system for monitoring salmon behavior, using specific patterns to identify stress or welfare concerns.
=========================================================
Topic 4: Non-supervised and Self-learned Annotation for AI-based Video Analysis
Training artificial intelligence models for image and video analysis requires a huge amount of data, and while for certain applications like traffic and people, massive open source data sets are available that is not always the case. For a lot of industrial applications, the availability of suitable data sets is limited. In this project we will seek to implement tools necessary for training models with a reduced (or zero?) need for manual annotation by means of automated data annotation. The project is suitable for one motivated student with interests in developing skills within programming and artificial intelligence methods for image data.
4.1 Main Objectives of the Project
============================================================
Topic 5: AI-based re-identification of salmon in large net-pens
Monitoring the health of each fish in an industrial net-pen is an important step towards individually tailored fish farming, which could greatly reduce welfare problems in aquaculture facilities. This master thesis project aims to construct an algorithm that can distinguish individual salmon by analyzing a selected part of the fish. The project will involve deep-learning-based appearance embedding networks, detection/segmentation models, and classical computer vision techniques.
Main objectives of the project
The objective of this project is to choose one part of the salmon (a fin, the head or the fish body), and
construct a workflow that uses this salmon part to distinguish between different individuals. The tasks of the pre-project can/will include:
Anticipated Tasks
The following tasks are anticipated to be necessary for the completion of the pre-project and master’s thesis:
===========================================================
Topic 6: AI-based Construction of 3D Salmon Models Using Monocular Videos
3D reconstruction of salmon can provide valuable information about the shape, mass, swimming patterns, and even texture of the fish. Access to this information is important for monitoring the growth, welfare, and total biomass in the salmon cages, and can be used for salmon re-identification. It is well known that monocular 3D reconstruction is an ill-posed problem, however, recent works have shown that sophisticated constraints and powerful AI methods can produce impressive models of quadrupeds and birds. These advances, however, have not been generalized to fish. As such, this project will examine the possibilities of monocular 3D fish reconstruction.
6.1 Main Objectives of the Project
Most state-of-the-art monocular 3D reconstruction methods use RGB images together with segmentation masks as input, and perform some sort of minimization of the reprojection error to iteratively refine the current 3D model. The main objective of this project is to create a workflow that generates individual 3D salmon models, given a monocular video of a single salmon. The tasks of the pre-project can/will include:
During the master's project in Spring 2024, the method can be extended and refined. Possible options include:
6.2 Anticipated Tasks
==========================================================
Topic 7: AI-based Pose Estimation and Tracking of Salmon from 3D Stereo Images using Computer Vision and Machine Learning
7.1 Introduction
Modern aquaculture relies on accurate measurements of fish size and mass for monitoring growth, welfare, and total biomass in each cage. This master's thesis project aims to develop an AI-based system to accurately estimate the size and mass of salmon from 3D stereo images. The project will involve the combination of deep learning for segmentation and detection to identify fish in a stereo camera setup, along with 3D reconstruction techniques for size estimation. This topic is well-suited for students interested in robotic vision, 3D reconstruction, and deep learning, with the exact focus of the work adaptable to individual interests.
7.2 Main Objectives of the Project
The primary goal of the project is to generate accurate pose and size estimates of salmon from image material, potentially using data from both experimental-lab and production facilities. The tasks for the pre-project may include:
During the master's project in Spring 2025, the method can be extended and refined. Possible directions for further development include:
7.3 Anticipated Tasks
=================================================
(This is a project in cooperation with Zeabuz)
Deep learning is increasingly playing a central role in perception systems for autonomous vehicles. Deep learning has potential to handle more complex data and scenarios than traditional model-based methods. On the other hand, a prerequisite for deep learning is large data sets for training. In the automotive industry this has been supported by systematic collection of data from large fleets of cars. In the maritime industry such data sets are currently lacking.
Zeabuz, a spinoff from NTNU founded in 2019, delivers maritime autonomy solutions for multiple maritime segments. Together with Torghatten, they established the world’s first commercial route with an autonomous urban passenger ferry in Stockholm in 2023, and delivers autonomy solutions in the workboat segment in the Singapore harbor area. With a rising number of vessels operating in various environments, the need to collect data, use it for training deep learning models and deploy them to vessels is increasingly important.
The goal of this project is to develop software for automatic collection of maritime sensor data, and to explore how these data can be used for training tracking pipelines based on deep learning. The project will make use of two key ideas.
First, since data from camera, radar and other sensors accumulate rapidly, the recording should be limited to relevant data in situations of interest. To achieve this, one may for instance use a radar-based tracker to zoom in on the relevant part of a camera image, and only record when there is an active radar tracking. In this way, one avoids recording terabyte after terabyte with empty sea.
Second, by comparing the camera data with radar data it is possible to do semi-supervised training.
Proposed tasks for the 5th year project
The goal of the specialization project is to make a pipeline for efficient data gathering, to be used in conjunction with the situational awareness system of Zeabuz. This entails the following tasks:
Proposed tasks for the master thesis
The goal of the MSc thesis is to use data recorded in this manner to develop and train tracking methods that make heavy use of deep learning. Tasks for the MSc thesis include:
References
Hangerhagen, P.: “A Benchmark Radar-Based Dataset from the Canal in Trondheim”, MSc thesis, NTNU, 2024
Meinhardt, T., Kirillov, A., Leal-Taixé, L. and Feichtenhofer, C.: “TrackFormer: Multi-Object Tracking with Transformers”, CVPR, 2022.
(This is a challenging thesis project for students with a good background in Deep Learning and Reinforcement Learning, and a basic understanding of Automatic Control – or vice versa.)
Background and MotivationAutonomous surface vessels (ASVs) are increasingly utilized in marine monitoring, environmental surveillance, and harbor logistics. For small-scale ASVs operating in dynamic and partially known environments, safe and efficient navigation remains a core technical challenge. Traditional planning and control frameworks—while effective under stable conditions—struggle to adapt to unexpected obstacles or rapidly changing contexts.
Recent work by Henrik.Fjellheim. (NTNU, 2023) demonstrated a foundational approach for short-term trajectory planning using Reinforcement Learning (RL) in such scenarios. Building on this foundation, the proposed MSc project will develop a more advanced hybrid planning and control framework that integrates Reinforcement Learning with Model Predictive Control (MPC). The combination aims to leverage the adaptive capabilities of RL with the real-time optimization and robustness features of MPC.
ObjectivesThe primary goal of this thesis is to design, implement, and evaluate a motion planning and control agent that enables robust, adaptive trajectory generation for small-scale ASVs. The system should be capable of:1. Interpreting the surrounding environment from sensory data and encoding it into a structured representation.2. Generating feasible actuator input sequences using a learning-based Action Sequence Generator (ASG).3. Validating and refining these sequences using MPC to ensure dynamic feasibility and optimality.4. Operating effectively in cluttered and uncertain environments, avoiding obstacles while progressing toward a goal state.
Envisaged Approach:The thesis will employ a time-discrete motion model for the ASV and simulate a range of environmental conditions. The RL agent will learn to map environment representations and long-term plans (e.g., paths or waypoints) to short-horizon control sequences. These sequences will then be evaluated by an MPC module that ensures:- Compliance with the ASV’s dynamic constraints.- Adherence to a cost function balancing goal proximity, energy efficiency, and smoothness of motion.- Safety through rigorous collision-checking using geometric and learned representations of obstacles.
Environment representations will be explored in various formats—including occupancy grids, circograms, and neural embeddings—to determine the most effective encoding for learning. The RL agent will be trained through simulated interaction, with the MPC serving as both a trajectory validator and a refinement engine.
Expected Contributions- A hybrid planning architecture integrating RL and MPC in the context of small ASVs.- Empirical analysis of different environmental representations and their impact on learning efficiency and navigation performance.- A simulation-based evaluation benchmarked against baseline planners and controllers.
Prerequisites and ToolsThe candidate should have prior experience with Python, basic control theory, and familiarity with machine learning (preferably RL). Tools likely to be used include PyTorch or TensorFlow, ROS2.
Referece:Background and MotivationAutonomous surface vessels (ASVs) are increasingly utilized in marine monitoring, environmental surveillance, and harbor logistics. For small-scale ASVs operating in dynamic and partially known environments, safe and efficient navigation remains a core technical challenge. Traditional planning and control frameworks—while effective under stable conditions—struggle to adapt to unexpected obstacles or rapidly changing contexts.
Recent work by H.F. (2023) demonstrated a foundational approach for short-term trajectory planning using Reinforcement Learning (RL) in such scenarios. Building on this foundation, the proposed MSc project will develop a more advanced hybrid planning and control framework that integrates Reinforcement Learning with Model Predictive Control (MPC). The combination aims to leverage the adaptive capabilities of RL with the real-time optimization and robustness features of MPC.
MethodologyThe thesis will employ a time-discrete motion model for the ASV and simulate a range of environmental conditions. The RL agent will learn to map environment representations and long-term plans (e.g., paths or waypoints) to short-horizon control sequences. These sequences will then be evaluated by an MPC module that ensures:- Compliance with the ASV’s dynamic constraints.- Adherence to a cost function balancing goal proximity, energy efficiency, and smoothness of motion.- Safety through rigorous collision-checking using geometric and learned representations of obstacles.
References:Short-Term Trajectory Planning for a Non-Holonomic Robot Car: Utilizing Reinforcement Learning in conjunction with a Predefined Vehicle Model
Henrik Fjellheim, 2023https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3097369?show=full
(This project is a cooperation with the company Fugro)
Background: Although the situation awareness system works well in good sunny weather, this is just the best-case scenario, not so common in offshore maritime practice. Adverse weather conditions such as fog, heavy rain, large waves, and strong wind can severely limit the functionality of sensors and cameras. For instance, foggy weather means that the vessel in control and other vessels have limited visibility; heavy rain means the visibility of the camera will be limited, and radar and lidar will suffer interference; heavy waves mean that the vessel's maneuverability will be compromised. While the performance of some sensors can be tuned to tolerate the impact of the weather to a certain degree, the other sensors become unusable. Thus, the onboard situational awareness system, including corresponding target detection, tracking, and sensor fusion algorithms, must adapt to the weather conditions to guarantee a certain minimal level of performance.
Scope: The project's first goal is to develop weather assessment methods based on available weather forecasts and sensory measurements from onboard sensors, including cameras, LIDAR, radar, and IMU. The second goal is to introduce the adaptivity mechanism to the target tracking and sensor fusion algorithms to optimize the overall system's performance based on the available weather estimation. Most of the project is assumed to be developed based on the simulated data. However, a few data-gathering campaigns using Fugro's test USV will be conducted to gather actual sensory datasets in different weather conditions to evaluate the performance of the developed algorithms.
Prerequisites: The candidate should have had courses in machine learning and/or computer vision and strong programming skills in Python and/or C++.
References: Zhang, Yuxiao, et al. "Perception and sensing for autonomous vehicles under adverse weather conditions: A survey." ISPRS Journal of Photogrammetry and Remote Sensing 196 (2023): 146-177.
Hakim, Arif Luqman, and Ristiana Dewi. "Automatic rain detection system based on digital images of CCTV cameras using convolutional neural network method." IOP Conference Series: Earth and Environmental Science. Vol. 893. No. 1. IOP Publishing, 2021.
With recent advances in digital technologies, the maritime industry is moving towards utilizing digital twins for ship inspection and maintenance. Combined with increased use of robotics, e.g. flying drones and crawlers, for inspection, we are exploring the potential of creating 3D digital twins from the data collected by the robots.
Ship hulls/tanks are built using known structural component types. The overall design of ship hulls is standardized, so the topological rules are known, i.e. how structural components relate to each other. This can be utilized for creating a 3D digital twin.
A goal is to develop a method for creating digital twins of ship hulls/tanks, enriched with semantic information of the structural components, using primarily visual data.
Lidar point clouds of the ship tanks, and camera positions for the visual data, are available in most cases, and can be used if needed. But an ideal method would rely mainly on visual data.
A particular challenge for ship tanks is that they usually have no ambient light. The only light source is the light that is brought along by the robot. Therefore, the illumination will change when the robot moves around, and distant objects will typically be darker in the images/videos.
Topic 1 – Semantic/Instance segmentation of ship hull structural components
Suggested tasks:
Topic 2 – 3D representation of the vessel hull based on monocular visual data
There is also the possibility to apply for a summer job at/with DNV in 2025 that is connected to this topic
Robust object detection is critical for enabling Unmanned Surface Vessels (USV) to perceive and understand its environment. A prerequisite for training such models are varied and large annotated training datasets describing the scenario the USV will operate in. However, annotating sufficient maritime data is extremely expensive and time-consuming.
Traditional random sampling approaches for training data selection can lead to redundant or uninformative samples, resulting in inefficient model training. Effectively selecting the most informative samples for annotation can drastically minimize annotation effort.
This thesis aims to investigate and develop effective sample selection strategies for maritime object detection in the context of the Maritime Robotics Otter USV.
Research Objectives:
Unmanned Surface Vessels (USV) rely on robust perception for safe navigation and obstacle avoidance. Robust perception requires camera-based instance segmentation for detecting and identifying objects in the environment. However, most existing methods utilize only camera data for detection.
This thesis aims to explore fusing LIDAR and camera data for 2d instance segmentation in the camera view without requiring expensive 3d annotations of the LIDAR point cloud. It will focus on near-shore detection, using data collected from Maritime Robotics SeaSight mounted on the Otter USV.
1. Develop a detection model to integrate LiDAR and camera data for 2d instance segmentation in the image frame. A starting point (and reference baseline) is reprojecting theLIDAR point cloud to the image frame and inputted to a YOLO instance segmentation model by concatenation with an RGB image.
2. Evaluate various fusion techniques for improved detection to determine how LIDAR data can effectively be integrated into a camera-based instance segmentation model.
3. Analyze the performance improvements achieved by the fusion approach compared to single-sensor detection.
4. Integrate the proposed model into the Maritime Robotics SeaSight system for experimental validation of the proposed model.
(THis is a cooperation project with NINA, the Norwegian Institute for Nature Research.
Problem description
Knowledge of the yearly spawning population size is important for the management of wild Atlantic salmon in Norwegian rivers. In the Tana watercourse, monitoring of the numbers, species and size of migrating fish is currently done by a combination of manual sonar and video-analyses. Sonar can be used for fish length measurements and has the advantage of independent of the light conditions, but cannot separate between fish species. The video data can be used for species recognition, but not for length measurements. The methods are thus complimentary. These manual analyses are time-consuming and results are not available until months after the recordings were made.
We would like to offer 1-2 master projects on automation of fish recognition, species recognition and fish length measurement using the sonar and video recordings. The video monitoring is done 24/7 using four underwater cameras per monitoring site, with a frame rate of 5-6 fps. We have large amounts of video recordings that could be used for training an AI to primarily recognize fish and capture images of this, secondarily to identify the species, and thirdly to do fish tracking based on the frames with identified fish. We do also collect new data each summer, and can offer field work in relation to this. NINA has been involved with another project regarding fish species recognition
from underwater video (https://github.com/beuss-git/bachelor-oppgave-nina), this work could potentially be continued for the Tana case, but there are other AI methods that could be tried. One of the challenges to AI is the (varying) natural background and varying light conditions. The fish counting and length measurement is done by adaptive resolution imaging sonar. Several softwares are available for automation of this task, but none of them works with a satisfying accuracy and precision. We cooperate with LUKE in Finland, which has developed one such software (Fishtracker, https://github.com/lukefi-private/FishTracker). This software is still premature, but promising. However, development of this has stalled due to lack of funds. The software is written in Python, and has a great potential for improving the tracking capabilities as well as automation and integration with automated analyses of the concurrent video recordings. The ultimate goal would then be to have a tool that can process and publish fish migration counts to species and size class “on the fly”, which would be a great tool for many of our larger salmon rivers with adaptive catch management. We are open to suggestions and modifications of these tasks by the student and supervisor from NTNU.
Data
We have large amounts of sonar and video data from previous years, including manual length measurements and species recognition with time stamps. These data can be used for model training as well as for assessment of model performance. These data are readily available on NINA servers, and can be made available on external hard disks if desirable. In addition, we will continue monitoring in the summer of 2024 and 2025, allowing testing of “on the fly” counting and classification when the methods for such
are ready.
Risks / Challenges
There should be no risks involved with the project.
The Norwegian Institute for Nature Research (NINA) is Norway’s leading institution for applied ecological research, with broad-based expertise on the genetic, population, species, ecosystem and landscape level, in terrestrial, freshwater and coastal marine environments.
The Norwegian Institute for Nature Research, NINA, is as an independent foundation focusing on environmental research, emphasizing the interaction between human society, natural resources and biodiversity. NINA was established in 1988. The headquarters are located in Trondheim, with branches in Tromsø, Lillehammer, Bergen and Oslo. In addition, NINA owns and runs the aquatic research station for wild fish at Ims in Rogaland and the arctic fox breeding center at Oppdal.
NINA’s activities include research, environmental impact assessments, environmental monitoring, counselling and evaluation.
NINA’s scientists come from a wide range of disciplinary backgrounds that include
biologists, geographers, geneticists, social scientists, sociologists and more. We have a broad-based expertise on the genetic, population, species, ecosystem and landscape level, in terrestrial, freshwater and coastal marine ecosystems.
In-house contact person at NINA:
Name: Karl Øystein Gjelland
Email: karl.gjelland@nina.no
Students interested in the project should, however, contact the prospective supervisor, Prof R.Mester at IDI first.
Background: While the LIDAR sensor has proven its efficiency for collision avoidance and docking use cases in harbors and inland water areas, its sensitivity to adverse weather conditions such as fog and rain makes it less attractive for marine offshore applications. Within radar technologies, X-band radar (around 8–12 GHz) is often found on ships for navigation and collision avoidance, as it readily spots distant targets. However, this band falls short in close-proximity scenarios, such as working near offshore structures. The recent advancement of W-band radar (around 77–81 GHz) introduces a viable alternative that delivers faster update rates, provides centimeter-level precision with 1 km range, and all-weather reliability. However, better sensitivity and resolution introduce a higher noise level into the results of sensor fusion algorithms for target detection and tracking.
Scope: The main goal of this project is to design and test target detection and tracking algorithms that better match the performance of the W-band radar and fuse the target tracking results with other conventional situational awareness sensors such as cameras, AIS, and X-band radar. The project's results must be verified in real conditions using Fugro's test USV during dedicated field tests.
Prerequisites: The candidate should have knowledge of modern target detection, target tracking, and sensor fusion algorithms and strong programming skills in Python or C++ to be able to deploy the developed algorithms on the real test USV for field tests. A background in computer vision or radar signal processing may also be useful.
References: Jang, Hyesu, et al. "MOANA: Multi-Radar Dataset for Maritime Odometry and Autonomous Navigation Application." arXiv preprint arXiv:2412.03887 (2024).
This project aims to advance the integration of LiDAR and stereo camera data to improve perception capabilities in autonomous mobile systems. By developing AI-driven fusion techniques, the goal is to achieve more accurate environmental understanding, benefiting applications such as navigation, obstacle detection, and object recognition. The research can be applied to various platforms, including:
- AutoDocking26 Project: A collaborative initiative focusing on autonomous docking for small-scale boats like the BlueBoat, designed for hydrographic surveys and robotics development. [BlueBoat website](https://bluerobotics.com/store/boat/blueboat/blueboat/)
- milliAmpere2 Autonomous Ferry: A full-scale autonomous passenger ferry platform operating in Trondheim waters.
- NTNU Revolve Racing Cars: High-performance autonomous racing vehicles developed by NTNU's Revolve student group.
- AgileX LIMO Mini-Robot: A compact mobile robotic platform for indoor/outdoor research in navigation and perception. [LIMO website](https://global.agilex.ai/products/limo)
Alternatively, the research can be adapted to other existing platforms within student projects.
Detailed Exposé
Autonomous systems rely heavily on accurate perception to navigate and interact with their environments. Combining data from multiple sensors, such as LiDAR and stereo cameras, enhances environmental understanding by leveraging the strengths of each modality.
In Prof. Mester's research group, multiple ongoing MSc and PhD projects focus on applied AI for perception, spanning a wide range of real-world autonomy problems. Most of these efforts are conducted in cross-departmental cooperation, particularly involving the Department of Engineering Cybernetics (ITK). This interdisciplinary setup enables research that is both conceptually rigorous and practically relevant.
AI plays a central role in these efforts, enabling flexible, learned perception systems that adapt to changing environments and sensor modalities. This project builds directly on that foundation by targeting AI-driven fusion of LiDAR and stereo vision data.
This project description serves as an umbrella proposition that can be concretized in collaboration with specific ongoing initiatives — for example, NTNU’s maritime autonomy projects — depending on the student’s interests and available partnerships.
This project is to be performed in close cooperation with the company Jotun
Transformer architectures have revolutionized sequence modeling, but their potential for structured spatial reasoning under uncertainty is only beginning to be understood. This MSc project investigates how Transformer-based models can detect, infer, and explain partially occluded patterns in visual or symbolic data — a setting where key structural elements are deliberately hidden.
The focus is on the mathematical and information-theoretic underpinnings of how attention mechanisms manage uncertainty and resolve ambiguity. Students will explore how self-attention layers act as dynamic superposition decoders, selectively amplifying consistent hypotheses while suppressing noise. We will draw from tools such as mutual information, rate-distortion theory, and vector superposition coding to analyze how occluded signals are internally represented and recovered.
Possible directions include: • Designing synthetic occlusion tasks with known symmetry/compositionality. • Analyzing internal attention maps and activations to detect latent structure. • Theoretically bounding the conditions under which reconstruction is possible.
This project is ideal for students excited by deep representation learning, transformer interpretability, and rigorous mathematical modeling. It blends cutting-edge ML with foundational theory — aimed at those who want to not just use Transformers, but understand why they work.
This topic is a cooperation with Fugro
Background: The use of deep neural networks to generate point clouds is gaining traction due to the interest in self-driving vehicles. The possibility of deriving depth from RGB cameras lead to generation of point cloud with a direct connection to 2D images, with is advantage when compared to other modalities such as lidar and radar which don’t display color. Furthermore, being able to couple point cloud generate from a stereo pair to other modalities tend to increase the robustness of the system by providing redundancy.
However, the absolute majority of research to build point clouds from stereo RGB cameras are focused on acquisitions obtained in land. Moreover, in areas with easily identifiable features, such as road signs, vertical references, and objects of similar size (cars, pedestrians, traffic lights, road lanes).
The marine environment presents more challenging aspects, such as a lack of references and landmarks to define a vanishing point. Marine images tend to contain larger portions representing the sky, which is an unreliable landmark due to the lack of strong gradients, but mostly due to the distance of the range of cloud to other elements visible in the camera range. The water presents perennial features, it should be considered but not be the focus when calculating a dissimilarity map. The focus should be on the objects and obstacles that can led to collisions.
Scope: The goal is to introduce an attention mechanism in the neural network architecture to detect features of interest. The data for training will be based on simulated data but also on stereo detections obtained with sensors of other modalities which are capable of detecting point clouds. The project's results must be verified in real conditions using Fugro's test USV during dedicated field tests.
References: Guo, Meng-Hao, et al. "Attention mechanisms in computer vision: A survey." Computational visual media 8.3 (2022): 331-368.
(This is a project topic in cooperation with Zeabuz)
Over the past decade, NTNU has developed concepts and technology for self-driving – or autonomous – marine vessels. A significant outcome of this research is the construction of the prototype passenger ferries milliAmpere1 and milliAmpere2, which are world-leading platforms.
The milliAmpere1 is a half-scale prototype for an autonomous urban passenger ferry and milliAmpere2 is a full-scale ferry which can carry actual passengers, and which performed the world’s first trial operation with an autonomous urban passenger ferry in the Canal in Trondheim in September and October 2022: https://www.youtube.com/watch?v=j3v47HiJmos&list=PLc2vvxBHfBcoHvfcIRsFROmJzXhbJCvb5&index=3
In 2019, the spinoff company Zeabuz was created to commercialize NTNU’s research on autonomous urban passenger ferries. Together with the Norwegian transport company Torghatten, Zeabuz has establish the world’s first commercial route with an autonomous urban passenger ferry in Stockholm urban passenger ferry in Stockholm in June 2023.
Zeabuz is also developing technology for automation and autonomy in the workboat segment and are currently testing this technology on a new electric workboat developed by Yinson, intended for operation in the Singapore harbour area. This harbour is the second most trafficked harbour in the world, and has several thousand workboats such as tugs, crew-transfer vessels, pilot vessels and supply-vessels that are servicing the fleet of merchant ships visiting the harbour. To operate autonomously in this area, the autonomy system needs to interact smoothly with other traffic and comply with the traffic rules that apply.
Autonomous platforms operating in the maritime domain are typically equipped with a comprehensive sensor suite with multiple sensing modalities. Some of these sensors provide explicit depth data, such as lidar and radar, making them preferred sensors for maritime target tracking. In addition, vision-based sensors are also often equipped to enable classification of detected objects. These sensors lack explicit range data, making them less useful for tracking.
Several techniques have over the years been developed to extract range data from cameras, usually through some form of geometric estimation. Stereo cameras are the current gold standard but require twice the number of cameras, doubling both cost and the required data bandwidth. Geometric processing of monocular camera data has also been tried before but this requires precise calibration and highly accurate navigation data.
In more recent years, neural networks have been developed for monocular depth estimation. Current state-of-the-art models can provide both relative and absolute depth measurements with a high degree of consistency. This project aims to investigate how these models can be used for situational awareness in the maritime domain.
This list is not exhaustive, and the focus can be shaped to fit the candidate’s preference and experience. Real-world experiments are expected and will be facilitated by Zeabuz and NTNU.
This task is comprehensive and requires a candidate that is motivated to work in the field of situational awareness for maritime vessels and autonomy. The candidate should be self-driven and structured. In return we offer close supervision with bi-weekly meetings and additional follow-up when needed.The project is a unique combination of theory and practice, combining development and implementation of algorithms that can be tested on real-world data, either recorded or live. The candidate should have completed TTK4250 Sensor Fusion and have a strong background in deep learning and computer vision.
The rise of video game streaming platforms like YouTube and Twitch has led to an explosion of gaming-related video content. However, categorising and analysing this vast content manually is impractical. This project proposes the development of an automated system that uses computer vision and NLP techniques to identify, classify, and categorize video game content in streaming videos.
This project focuses on developing an ML-based system capable of segmenting and classifying gaming videos into different emotional and thematic categories. The goal is to automatically assign content/narrative percentages to various content types present in a live streamed gaming video — for instance, identifying that a stream consists of 10% exploration, 30% combat, 20% high-tension moments, and 40% calm narrative or idle periods etc. This goes beyond traditional object or HUD detection, aiming instead to capture the mood, pacing, and thematic shifts within gaming content.
Objectives:
Data collection: gather a large dataset of gaming video clips from YouTube and Twitch across multiple popular video games and genres.
Feature extraction: develop a deep learning model to identify sequential content through feature extraction (check the following references).
Employ audio/NLP analysis to gauge the emotional tone of each segment.
Multimodal model: integrate audio-visual fusion models to combine insights from both modalities for richer understanding.
Temporal analysis: implement sequence models (e.g., LSTMs or Transformers) to ensure smooth and coherent classification across video timelines.
Content summarisation: aggregate segment classifications to generate an overall content breakdown (e.g. 10% exploration, 30% combat, 20% high-tension moments, and 40% calm narrative or idle periods.)
Evaluation: evaluate model performance using standard metrics like accuracy, precision, recall, and F1-score, and test generalisation on unseen games or streams.
Conduct qualitative evaluations comparing automated summaries to human annotations.
Technical Considerations:
Use video processing libraries to extract frames at regular intervals.
Develop semi-automated tools to assist in the manual annotation process.
Fine-tune pre-trained models, apply multi-label classification architectures, and use temporal models for sequence consistency.
Incorporate sequential models to capture temporal patterns across video sequences.
Use transfer learning to handle limited labelled data for niche games.
Expected Outcomes:
A recommender system based on identified video game narrative
Desired Candidate Skills:
Proficiency in Python, deep learning frameworks (TensorFlow/PyTorch).
Experience in computer vision and NLP/audio processing.
Interest in gaming and understanding of different gaming genres.
References:
Yeung, S., Russakovsky, O., Jin, N., Andriluka, M., Mori, G., & Fei-Fei, L. (2018). Every moment counts: Dense detailed labeling of actions in complex videos. International Journal of Computer Vision, 126, 375-389.
Anderson, P., He, X., Buehler, C., Teney, D., Johnson, M., Gould, S., & Zhang, L. (2018). Bottom-up and top-down attention for image captioning and visual question answering. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 6077-6086).
Yu, H., Wang, J., Huang, Z., Yang, Y., & Xu, W. (2016). Video paragraph captioning using hierarchical recurrent neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 4584-4593).
Schwenzow, J., Hartmann, J., Schikowsky, A., & Heitmann, M. (2021). Understanding videos at scale: How to extract insights for business research. Journal of Business Research, 123, 367-379.
Haroon, M., Wojcieszak, M., Chhabra, A., Liu, X., Mohapatra, P., & Shafiq, Z. (2023). Auditing YouTube’s recommendation system for ideologically congenial, extreme, and problematic recommendations. Proceedings of the national academy of sciences, 120(50), e2213020120.
Yakaew, A., Dailey, M. N., & Racharak, T. (2021, February). Multimodal Sentiment Analysis on Video Streams using Lightweight Deep Neural Networks. In ICPRAM (pp. 442-451).
Karjee, J., Kakwani, K. R., Anand, K., & Naik, P. (2024, January). Lightweight Multimodal Fusion Computing Model for Emotional Streaming in Edge Platform. In 2024 IEEE 21st Consumer Communications & Networking Conference (CCNC) (pp. 419-424). IEEE.
Lightweight Models for Emotional Analysis in Video. https://arxiv.org/abs/2503.10530
Introduction
Amidst a burgeoning aging populace and surging demand for homecare (hjemmetjenester) services, there emerges an urgent call for innovation and optimization in service delivery. Decentralized technology presents distinct advantages, such as user-centric data ownership, accessibility, heightened privacy, security, and transparency. These attributes hold the potential to significantly enhance the efficiency and efficacy of municipal homecare services. The proposal aims to explore how decentralized technology could transform and enhance homecare services administered by Trondheim municipality, heralding a paradigmatic shift in service provision.
Central to the project's focus is the introduction of a decentralized, user-owned health wallet platform to tackle the myriad challenges facing contemporary homecare services. This pioneering platform empowers individuals to assume control over their health data, enabling secure storage, management, and sharing of medical information according to their preferences. Against the backdrop of mounting concerns surrounding privacy breaches and data mishandling, this initiative offers a compelling alternative by reinstating ownership and oversight of sensitive health data to patients, thereby fortifying security and privacy protocols. Moreover, it promises to foster transparency and bolster patient autonomy, fostering active participation in their healthcare journey. Note that this isn't just about data – it's also about empowerment of patients!
Furthermore, the envisioned platform holds the potential to revolutionize not only homecare services but also broader healthcare provision by facilitating seamless data exchange among patients, healthcare providers, and other stakeholders. Such enhanced connectivity ultimately promises improved outcomes and a more patient-centric healthcare ethos.
Tentative tasks
Task: Develop an app for mobile devices, that can be mounted in public transportation like busses, can access the camera of the device as well as relatively cheap and accurate positioning equipment with CPos corrections (cm accuracy) and have AI models for assessing and geo-referencing the condition of all road objects visible from the road (one application, other applications could be to create and update HD-maps, match real-time images to a reference for back-up localisation, collect data for neural rendering etc.).
Collaboration: the project will be a collaboration between the national road and mapping authorities, SINTEF and several counties / municipalities. Huge innovation potential.
Supervisors: Gabriel Kiss kiss@ntnu.no Frank Lindseth frankl@ntnu.no (COMP/IDI)
In an era marked by increasing digital transactions and online interactions, ensuring the security and integrity of personal identities has become paramount. Traditional methods of identity verification, such as passwords and biometrics, are often susceptible to fraud and exploitation. However, the integration of Artificial Intelligence (AI) offers a promising avenue for strengthening identity security measures. By harnessing AI algorithms for identity verification, organizations can enhance accuracy, efficiency, and resilience against fraudulent activities. This proposal seeks to explore the implementation of AI-driven identity security systems to fortify the protection of individuals' personal information and prevent identity theft.
Tentative Tasks:
Investigate the use of Generative AI, such as Large Language Models (LLMs), to configure static analysis tools (such as SonarQube, PMD, etc.), with particular focus on defining customized coding rules. These tools are very useful for discovering software faults, but they are difficult to configure and to customize. This project wants to understand if and how LLMs can help with this task.
Project Description
Static code analysis tools [1] (often referred to as “linters”), such as for example SonarQube [2], PMD [3], or SpotBugs [4], are widely used to identify common bugs and mistakes in programming. They are based on identifying coding patterns that are known to introduce faults or vulnerabilities in the code.
While extensive coding rules exists, such as SEI CERT Coding Standards for Java [5] for security, or MISRA C/C++ [6] for safety, these rules evolve with the discovery of new bugs and vulnerabilities, or with the introduction of new features of programming languages. Further, developers may want to define customized rules to cover internal patterns or coding standards that are adopted by their company.
Most of these tools can be customized with new rules, but the process is typically quite cumbersome (e.g., [7] [8]). Generative AI (GenAI) models such as Large Language Models (LLMs) has shown disruptive performance on tasks such as text processing and code generation, and research on the use of GenAI for software engineering tasks is emerging. This project aims to investigate how LLMs can help in configuring static analysis tools.
The idea is to use LLMs translate rules specified in natural language, to a configuration of the static analysis tool. Data will be obtained from the hundreds of rules already implemented in open source static analysis tools, such as PMD.
This work proposal involves:
The long-term research objective linked to this activity is to simplify the definition and verification of coding rules, through the use of GenAI.
Recommended skills
Critical systems are those systems whose failure may cause severe harm or financial consequences, such as for example systems within the railway or aviation domain. Those systems must undergo a rigorous verification,validation, and certification process, which involves very time-consuming activities.
Because those systems cannot be easily tested, part of the process involves early evaluation through different kind of models. For example, Fault Trees (FT) [1] are extensively used to understand the possible failure of system components and the consequences at system level. However, building those models is very time consuming and requires specialized skills. Furthermore, those models need to be updated when some parts of the system design change. This project aims to exploit the potential of generative artificial intelligence (such as LLMs) to automate the construction of models of critical systems, such as Fault Trees (FT) and Stochastic Petri Nets (SPN), from a higher-level textual or semi-structured description.
The project will be based on the Agentic AI paradigm, where multiple LLM-based agents focus on specific tasks, and collectively contribute to a solution [1][2].
Because research in using LLMs in critical systems is in its infancy, part of the pre-project will also focus on building a representative dataset for this kind of task, based on material from the literature available online. Building this dataset has a strong potential to influence the future research on this topic.
Tasks
Semi-formal models (e.g., UML diagrams) are used in different tasks of system and software engineering, for example for documenting the software architecture. While modeling tasks are a creative effort, they also require much manual effort and they are typically error prone and difficult to be maintained. Also, with the increased use of AI, it becomes increasingly difficult to understand and interpret the sofware architecture of generated code. This project aims to exploit the potential of generative artificial intelligence (such as LLMs) to automate the extraction of software architecture models, based on software repositories on GitHub.
Response technology (response systems) allows teachers to ask questions to large groups of students and get aggregated and useful answers to guide the lecture. Most of the existing systems require preparing the questions in advance and offer little to no flexibility in asking ad hoc questions or even using the results from a question as the basis for a follow-up question.
The primary aim of this project is to design and implement an agile question generation approach that analyzes student open-text responses and produces contextually relevant follow-up questions during interactive lectures. While existing question-generation solutions focus on structured content, using open-text student responses for question generation in real-time remains challenging. Additionally, there is a lack of empirical evaluation of these systems in classroom environments.
With advances in artificial intelligence and natural language processing, automated question generation could be a promising technique for enhancing interactive learning environments. The effectiveness of the proposed solution could be evaluated through user studies, assessing its impact on student engagement, learning outcomes, and teaching adaptability. Teachers can dynamically adapt their teaching strategies by generating meaningful follow-up questions based on student responses, probing deeper into students' understanding, and fostering productive discussions.
While the project can be assigned to a single student, it is recommended that a pair of students work on it.
Enterprise Modeling has been defined as the art of externalizing enterprise knowledge, i.e., representing the core knowledge of the enterprise. Although useful in product design and systems development, for modeling and model-based approaches to have a more profound effect, a shift in modeling approaches and methodologies is necessary. Modeling should provide powerful services for capturing work-centric, work-supporting and generative knowledge, for preserving context and ensuring reuse. An approach to this is Active Knowledge Modeling (AKM). The AKM technology is about discovering, externalizing, expressing, representing, sharing, exploring, configuring, activating, growing and managing enterprise knowledge.
AKM is supported through an open source modeling product Mimris, available on the web Mimris Modeling App
AI is used to support both the development of modelling notations and enterprsie models, and how to use AI to develop the modeling environment, and support modelling is an important aspect
The task relates to the development of an AKM-solution for a specific problem domain, and evaluating this from the point of view of usability and usefulness.
The focus of this thesis is to develop an Artificial Intelligence based system to help the students learn mathematical concepts while playing educational games. One of the ways to provide help is to find out the difficult moments during the interaction and then supporting the students when they are faced with such moments. The challenging aspect of such projects is the “cold start problem”. We need to know in advance how to detect the difficult moments for individual students. Solving this problem will be a key aspect of this thesis
Thesis DescriptionIn a first step, the student(s) will design and implement the feedback tool using the wearable sensors. Afterwards, they will conduct a user study in order to test the usability of the system with a number of students. Once the usability of the system is established (with the last changes in the system), the student(s) will conduct a larger user study to evaluate the effectiveness of the system. Finally, the candidate(s) will analyse the collected data and write up his/her thesis.
RequirementsThe ideal candidate will have a background in system design and basic machine learning. Solid programming skills and an interest in hands-on development and experimentation is also a requirement.Programming skills: Python/Java.
The primary objective is developing and demonstrating an AI Assisted Modelling App, showing how AI could be used as an assistant for Modellers.
The workplaces produced will demonstrate AI collaboration and innovation principles and methodologies. AI Assisted Modelling implies intelligent user- and AI agent-driven balancing of properties, capabilities, qualities and services, reducing errors and change management, and cutting calendar times and costs by factors.
The secondary objectives are:
The web-based Modeling Platform has being implemented in an Equinor Accelerator project and will support the tasks to be performed. A Demonstrator of AI Assisted workplaces and capabilities extending the capabilities of the Mimiris Modelling Platform and recent digitisation approaches, such as Intelligent Agents and Digital Twins, will be implemented in demonstrators.
We will conduct this work with the Customer (Company): KAVCA AS.
(AI-SECRETT) The need for sustainable transitions requires all sectors to enhance their competencies and skills to achieve the social, economic and environmental transition. This could be facilitated with competencies related to AI and creativity. This project will focus on identifying the competencies related to AI and creativity that would help towards sustainable transitions and designing a solution that could help people enhance their skills and competences. The tasks will include:
The outcome could be a means of supporting the co-design of learning activities related to AI and creativity. This task is related to the European project AI-SECRETT.
The green shift is high on every executive’s agenda, and with good reason. The urgency of the climate crisis and associated transition to a sustainable society changes the way firms create, capture, and deliver value. Shifting the very fabric of today's business landscape. Firms must now deliver on a triple bottom line (environmental, social, and economic) and not only meet today's needs from customers and shareholders, but also future generations' needs and opportunities for value creation. A strategic response is required, and firms must make structural changes to accommodate a fully sustainable business model (SBM). Research suggests that firms that manage and mitigate their exposure to climate-change risks while seeking new opportunities for sustainable value creation will generate a competitive advantage over rivals in a carbon-constrained future. However, transitioning towards a SBM is challenging and companies often lack the necessary data and insight to make correct and effective business decision. Artificial Intelligence (AI) offers a possible solution by establishing a basis for data-driven and fact-based decision making. This makes it easier for firms to take a systems perspective, quantify impacts, and reduce the complexity of the sustainable transition. Although real and theorized examples of AI enabling SBMs exist, a comprehensive understanding of the relationship between AI and SBM is still missing, leaving a gap in our understanding of the underlying mechanisms and inhibiting firms’ ability to accelerate their sustainable transition. Thus, this project aims to take stock of current knowledge by studying the following research questions:
RQ1: What do we know about the relationship between AI and SBM? RQ1.1: How can companies leverage AI for SBM? RQ1.2: How can the relationship between AI, SBM and competitive performance be conceptualized?
Even though oceans are very important for human life and societies, we have very little understanding of marine ecosystems which are very complex systems. Ocean observatories and other underwater monitoring systems provide data streams that cover physical, chemical and biological ocean properties.
Marine animals are very vocal; many invertebrates, fishes, and nearly all marine mammal species produce sounds. Underwater acoustic data provide scientific insights in a broad range of fields including animal vocalizations (biophony) and anthropogenic noise (anthrophony) in the marine environment. However, underwater sounds are very challenging to identify and classify due to the variability of sound events and conditions under which they were recorded. This creates interesting research challenges to deal with this variation, sparsity and noise.
This project aims to develop effective classification and analysis models of large acoustic data streams and additional data from ocean observatories. The student(s) must be able to work with large amounts of data and willing to get familiar with transformer models.
The project will run in collaboration with the Department of Biology at NTNU. It will be possible to continue the work previous master's students have done.
Open text questions allow students to answer without being influenced by predefined options and thus eliminating some causes for bias and guessing.
The primary aim of this project is to develop an intelligent solution that allows a teacher to ask a knowledge related open-text question and get an aggregated overview that indicates with a certain level of confidence what percentage of students got it right, partially right, partially wrong, wrong etc. This will allow the teacher to offer adaptive feedback in order to clarify any misunderstanding. The project could be designed to provide a real-time dashboard for teachers, offering insights into student responses and highlighting common misconceptions that may require further explanation, identifying knowledge gaps, and adjusting lectures dynamically.
The implementation of natural language processing and artificial intelligence advancements can assess responses based on correctness, relevance, coherence, and depth of understanding. The dataset collected from student responses during lectures is expected to vary from a few tens to a few to hundreds of responses per question. The proposed solution’s effectiveness can be tested in real lecture environments, ensuring that it meets the needs of teachers and students. Ultimately, this research will contribute to advancing AI-driven education technologies, demonstrating how automation and intelligent feedback mechanisms can enhance teaching effectiveness and student learning experiences.
This project is proposed by Vend, a Nordic Marketplaces company building and powering leading Nordic marketplaces such as FINN.no, Blocket.se, and more.
The project will be supervised by the experts in the company's security engineering team.
The security engineering team is responsible for application security and cloud security across the organization. The team works with vulnerability management, secure development practices, and security tooling across a large portfolio of products and services.
Security engineers in the team regularly deal with high volumes of findings from automated scanning tools (SAST, SCA, DAST, cloud posture management, bug bounty programs) and coordinate with development teams to prioritize and remediate vulnerabilities. This project grows directly from the challenges the team faces daily.
Problem statements
Modern software organizations generate a high volume of security findings from diverse sources: static analysis (SAST), software composition analysis (SCA), dynamic testing (DAST), cloud security posture tools, and external sources like bug bounty programs. The core challenge facing security teams is not finding vulnerabilities. It is processing, prioritizing, and verifying them efficiently.
Security engineers face three compounding problems:
• Triage overload. Automated security scanners produce large volumes of findings, many of which are false positives or low-impact issues. Manually triaging these requires deep contextual understanding of the codebase, its deployment environment, and the organization's threat model. A single scanning run can produce hundreds of findings, and a security team may manage dozens of applications.
• Verification gap. Even when a finding is confirmed as a real vulnerability, verifying its exploitability remains manual and time-consuming. Is the vulnerable code path actually reachable? Can it be exploited given the application's specific configuration and deployment? Today this often requires a security engineer to manually trace data flows, construct a proof-of-concept, or perform targeted penetration testing.
• Knowledge bottleneck. Security expertise is scarce and unevenly distributed. Most development teams lack the depth of knowledge needed to assess whether a reported vulnerability is critical or irrelevant in their context. Security teams become a bottleneck when every finding requires their manual review.
Recent advances in large language models (LLMs) and agentic AI systems open new possibilities for augmenting security engineers across this workflow. This thesis project explores how AI can be applied to one or more stages of the security engineering pipeline:
• Vulnerability detection: Using LLMs to identify security issues in source code, infrastructure-as-code, or cloud configurations. This includes exploring whether LLMs can catch classes of vulnerabilities that rule-based static analysis tools miss, or reduce false positive rates by understanding code context more deeply.
• Triage and prioritization: Applying AI to contextualize and rank security findings from existing tools. This could involve training or prompting models to assess severity based on code context, deployment configuration, data flow analysis, and historical triage decisions made by security engineers.
• Automated verification: Using LLM-based agents to reason about exploitability, trace attack paths through code, or generate proof-of-concept exploits for reported vulnerabilities in controlled environments. This is the most ambitious direction, building on recent research in autonomous security agents and LLM-driven penetration testing.
Project focuses and tasks
The student(s) will select a specific focus area in collaboration with the supervisors and company contact during the project assignment phase. The expected outcomes are:
• A systematic evaluation of AI-based approaches for the chosen security engineering task.
• A working prototype or pipeline demonstrating the approach on realistic data.
• An empirical analysis comparing AI-augmented results against existing baselines (manual triage, traditional tools, or both), with honest assessment of where AI adds value and where it falls short.
The scientific challenge lies in understanding the capabilities and limitations of current AI models for security-critical reasoning: where they add genuine value, where they hallucinate or miss real issues, and how to design human-AI workflows that improve security outcomes without creating false confidence.
The practical challenge is building something that could realistically integrate into a security engineering workflow and measurably reduce time-to-triage or improve detection and verification accuracy.
Required background
Basic knowledge of AI; Some knowledge of software security, preferably having taken the software security and data privacy course (TDT4237); Hands-on experience in developing prototypes for demonstration purposes.
This thesis will be carried out in collaboration with E.A. Smith AS and will be co-supervised by the company.
This thesis can be conducted by one or two students.
E.A. Smith AS is a long-established Norwegian trading company with strong positions in steel, metals, building materials, reinforcement solutions, and tools. Within its steel business, pricing decisions are commercially important and often complex. Today, sales representatives have a degree of flexibility in price setting, but the current process is not fully data-driven and may lead to inconsistent pricing and margin erosion. The company therefore wants to explore how artificial intelligence and machine learning can be used to support more robust and profitable pricing decisions in the sales process.
The thesis will focus on a specific business area called “Steel 10” in the B2B segment. This area has already been identified as a relevant and high-impact domain for pricing improvement. Existing analyses indicate that Steel 10 has meaningful pricing complexity, significant margin variability, and strong business relevance. A previous proof-of-concept has also demonstrated that it is possible to generate actionable price and margin recommendations by combining historical sales data, product grouping, customer segmentation (clustering), and price and margin recommendation logic.
The main objective of this thesis is to design, develop, and evaluate an AI-based pricing recommendation solution for B2B steel sales. The solution should support sales representatives by generating recommended prices and margin ranges for specific product, customer, and location combinations.
The thesis should investigate both the analytical and technical aspects of such a solution, including data preparation, machine learning methodology, recommendation logic, evaluation, and possible user-facing presentation.
The student(s) should analyze the pricing problem in the Steel 10 B2B segment and define a clear problem statement for the thesis. This includes:
understanding the current pricing process and its limitations,
identifying the decision points where sales representatives need support,
defining what constitutes a useful and trustworthy price recommendation,
clarifying how such a solution could create business value.
The student(s) should work with historical transaction data, relevant master data made available by E.A. Smith as well as an external pricing data source. The work should include:
exploring data quality, coverage, and suitability for machine learning,
understanding seasonality, customer behavior, and pricing variability,
applying relevant filters and transformation logic,
creating a structured dataset for modeling and experimentation.
Relevant data may include order history, customer attributes, item attributes, item hierarchies, location data, and item-location cost data.
A central part of the thesis is to investigate how machine learning can be used to segment products and customers in a meaningful way.
The student(s) should explore methods for:
clustering products based on features such as aggregated margin, total cost, and number of orders,
clustering customers based on their purchasing behavior across product groups,
evaluating whether these segments provide a useful basis for differentiated pricing recommendations.
The thesis may build on unsupervised learning methods, such as clustering techniques, but students are encouraged to evaluate alternative methods where relevant.
Based on the segmentation and historical transaction patterns, the student(s) should design recommendation logic for price and margin guidance. This may include:
identifying relevant historical benchmarks for similar products and customers,
defining lower and upper bounds for acceptable margin or price,
generating a recommended price and/or recommended margin,
discussing how recommendations should vary across customer segments (clustering), product groups, or locations.
The thesis should also discuss explainability and how recommendations can be presented in a way that sales representatives can understand and trust.
The student(s) should develop a prototype that demonstrates the proposed solution. The prototype does not need to be production-ready, but it should be sufficient to test the concept in practice.
Possible prototype components include:
a data processing and modeling pipeline,
a recommendation engine,
an API or service layer,
a simple user interface or widget mock-up that presents recommendations to end users.
The prototype may be implemented using modern data science and software lengineering tools, depending on the students’ background and thesis scope.
The student(s) should evaluate the technical and business relevance of the solution. Possible evaluation dimensions include:
model quality and segmentation quality,
stability and robustness of recommendations,
usability and interpretability,
estimated business impact,
feasibility of integrating the solution into a broader enterprise architecture.
Where possible, the student(s) should compare the proposed recommendations with historical outcomes and discuss the potential value of improved pricing support.
The available project documentation describes a pricing architecture where historical data is transformed into analytical datasets, processed through staging and machine learning steps, and used to generate price recommendations. The documented solution includes data transformation logic, product and customer clustering, margin-based recommendation logic, and a widget-oriented presentation layer for sales users.
This gives the thesis a strong practical foundation. At the same time, the student(s) are expected to critically assess the existing assumptions, improve parts of the analytical approach where relevant, and propose a solution that is academically sound and practically useful.
The thesis is expected to deliver:
a clear problem formulation and literature-based framing,
an analysis of the business and data context,
a documented machine learning and recommendation approach,
a prototype demonstrating the proposed solution,
an evaluation of the results,
recommendations for further development and industrial use.
This thesis will contribute to the development of AI-supported pricing in an industrial B2B context. It combines data science, machine learning, business decision support, and applied software design. The work is expected to provide both academic value and practical insights into how pricing recommendations can be generated and presented to support better commercial decisions in the steel business.
The intersection of artificial intelligence (AI) and healthcare presents an opportunity to enhance medical diagnostics, improve patient outcomes, and streamline clinical workflows. In previous years, this project has focused on developing an AI-driven web application capable of analyzing medical data to support decision-making for doctors and medical students.
This year's thesis will focus on further developing the system by improving existing functionalities, expanding diagnostic capabilities, and refining user experience. Additionally, the project will explore new machine learning models to enhance accuracy and reliability in medical diagnostics.
Students will be expected to extend the existing system and contribute to one or more of the following areas:
By participating in this project, students will continue the development of the AI-driven assistant, refine existing modules, and test improvements through a structured user study.
Project Objective
The primary goal of this project is to design and implement a web-based AI-driven diagnostic assistant that aids doctors in creating accurate diagnoses and helps medical students sharpen their digital skills. This assistant will harness the power of image processing, quantitative data analysis, and natural language processing (NLP) to analyse medical data comprehensively.
The ideal candidates should have:
Expected Project Work Packages
This master's thesis is part of the newly established AI Centre for the Empowerment of Human Learning (AI LEARN) — one of six national AI centers in Norway. AI-LEARN focuses on the interaction between humans and AI, with an emphasis on developing human-centred infrastructures designed to enhance and empower human learning.
Supervisor: Boban Vesin
Suitable for: Two students
The intersection of artificial intelligence (AI) and healthcare presents an opportunity to enhance medical diagnostics, improve patient outcomes, and streamline clinical workflows. In previous years, this project has focused on developing an AI-driven web application that analyzes medical data to support decision-making by doctors and medical students.
The primary goal of this project is to design and implement a web-based AI-driven diagnostic assistant that aids doctors in creating accurate diagnoses and helps medical students sharpen their digital skills. This assistant will harness the power of image processing, quantitative data analysis, and natural language processing (NLP) to comprehensively analyse medical data.
Supervisors: Michail Giannakos, Boban Vesin
This thesis explores how learning data from systems such as the Norwegian student information system FS (Felles studentsystem), Canvas, and Sikt (Kunnskapssektorens tjenesteleverandør) can be integrated through xAPI and LTI, and how AI can help turn this data into meaningful insights for students, teachers, and administrators.
As educational institutions increasingly rely on multiple digital platforms, interoperability and data integration have become essential for enabling effective learning analytics. This thesis explores interoperability frameworks such as the Experience API (xAPI) and Learning Tools Interoperability (LTI) within educational ecosystems, with particular focus on systems such as FS, Canvas, and other data flows managed by Sikt. The overall goal is to investigate how learning-related data from different systems can be accessed, combined, and presented in a meaningful way.
In addition to data integration, the thesis may explore how artificial intelligence can make learning analytics more understandable and useful for stakeholders. For example, AI-based functionality may support automatic summarization of learner activity, natural-language explanations of dashboard data, identification of relevant patterns, or other forms of insight generation based on integrated learning data.
The thesis is therefore positioned at the intersection of system interoperability, learning analytics, and AI-supported educational technology. It offers an opportunity to work with real integration challenges while also exploring how modern AI techniques can improve the usability and value of learning data.
Thesis Description
To make the topic relevant to current developments, the project may also include an AI-related component. This may involve, for example, AI-generated summaries of learner activity, natural-language interaction with learning analytics dashboards, intelligent interpretation of cross-system data, or other forms of AI-supported insight generation.
The exact scope should be defined by the student in collaboration with the supervisors, based on interests, available data sources, and technical feasibility.
Candidates should be comfortable with software development and interested in educational technology and data science. Relevant skills may include:
This remains consistent with the original proposal’s emphasis on software development, analytics, and visualization.
Teachers make rapid and complex decisions while managing classrooms, responding to students, and delivering instruction. Understanding these cognitive processes is crucial for improving teacher training, classroom strategies, and AI-driven educational tools. Eye-tracking technology, combined with Artificial Intelligence (AI), offers a powerful approach to analyzing how teachers allocate visual attention and make instructional decisions in real time. This thesis aims to explore how AI-enhanced eye-tracking can be used to study teacher behavior, cognitive load, and decision-making patterns in educational settings. By leveraging AI to process and analyze eye-tracking data, the research seeks to uncover insights that can improve teacher training and optimize classroom dynamics. By integrating AI and eye-tracking, this study will provide valuable insights into teacher cognition and instructional decision-making. The findings could pave the way for more adaptive AI systems that support educators in real-time.
The ideal candidate will have a background in system design and basic machine learning. Solid programming skills and an interest in hands-on development and experimentation is also a requirement.
Programming skills: Python/Java.
Electroencephalography (EEG) is an electrophysiological monitoring method to measure electrical activity in the brain. From noninvasive small electrodes that are placed along the scalp, EEG record spontaneous electrical activity in our brain. Analyzing EEG signal data helps researchers to understand the cognitive process such as human emotions, perceptions, attentions and various behavioral processes.
Brain-computer interfaces (BCIs), motor imagery (MI), and virtual reality (VR) offer a unique opportunity to engage the brain in ways that traditional therapy cannot, and the integration of VR/AR combined with electroencephalograms (EEG) is shown to enhance nervous system recovery compared to traditional rehabilitation. In this study, the students will investigate suitable maching learning modells for EEG classification and combine motor-imagery EEG classifier with interactive VR environment that incorporates core gamification elements. Further, it is also interesting to investigate to what extent immersion plays a role in motor imagery.
The study is associated with Vislab and the needed sensors and the basic training on how to use them (VR headset, EEG equipment) are provided.
Over a long time, we have performed surveys of the development and maintenance of IT-systems in Norwegian organizations. Comparable data for important areas where also capture in collaboration with Rambøll IT I praksis investigations, analysing the data from the 2026 investigation. A focus area in the last years is the development and implementation of AI-solutions, with a focus on the implementation of AI in public sector. The assignment will be to analyze the quantitative and qualitative data from recent investigations. Together with a literature review, the survey investigations are expected to give us new knowledge about mechanisms affecting resource utilization related to information systems support in organizations in particular in connection to taking AI into use. The report should be written in English and is expected to form the basis for scientific publications
One of the key trends of modern computing systems is certainly the push for increased efficiency, with focus both on energy-to-solution and time-to-solution metrics. It is possible to push the boundaries of the hardware limitations to save on these metrics via approximate computing techniques.
Another research aspect that is somewhat orthogonal relates to the resilience that a software systems shows in the presence of faults, which may induced by external actors (e.g., radiations) or by a malign entity (targeted attacks).
Our key research questions is simple: are these two aspects completely orthogonal or are they correlated?
This project aims at expanding previous research in the field of mixing compiler-level precision tuning with fault-injection to observe software reliability metrics.
Project in collaboration with Leonardo Montecchi (ISSE, NTNU)
For your own safety, please consider the following expected requirements and you will not hurt yourself while working on this project.
Mandatory:
Desired:
As decentralized technologies mature, modern platforms increasingly combine Web3 components (blockchain, smart contracts, decentralized storage) with traditional Web2 services (APIs, databases, authentication layers). Self-Sovereign Identity (SSI) adds another architectural layer, enabling secure, user-controlled identity management through verifiable credentials and decentralized identifiers (DIDs). Validating such hybrid architectures early is essential for ensuring interoperability, performance, and security. This project explores the use of Simulink and System Composer for modeling and validating the architecture of a Web2–Web3–SSI integrated system.
Model a hybrid architecture combining Web2 services, a Web3 backend (e.g., blockchain nodes or smart contracts), and an SSI identity layer.
Identify key architectural concerns: authentication latency, credential issuance/verification flows, cross-system data integrity, and fault tolerance across distributed components.
Use Simulink/System Composer to simulate message flows, network delays, identity interactions, and multi-layer service dependencies.
Validate interoperability and detect architectural risks prior to implementation.
Select a representative hybrid use case (e.g., Web2 app using SSI for login and a blockchain for audit logging or credential storage).
Build subsystem models representing Web2 backend, SSI agent/wallet, verifiable credential issuer/verifier, blockchain nodes, and communication channels.
Simulate scenarios such as credential verification under load, node failures, network latency spikes, and inconsistent DID resolution.
Analyze and verify the architecture models against - requirement coverage, functional correctness, quality and performance attributes, requlatory complience, etc.
Expected Outcomes
A Simulink/System Composer architecture that visualizes interactions between Web2 components, Web3 infrastructure, and SSI identity flows.
Simulation and validation results identifying performance bottlenecks, failure points, and integration challenges.
Architectural insights and recommendations for designing robust hybrid Web2–Web3–SSI systems.
The last few years have seen an explosion in interest regarding the use of Artificial Intelligence and much talk about the potential business value. Nevertheless, there is significantly less talk about the challenge's organizations will face when implementing such solutions and how they should overcome these obstacles. Inhibiting factors are not only of a technological nature but also include organizational and human factors. This project will involve collecting and analyzing data in collaboration with the researchers from the Big Data Observatory (https://www.observatory.no).
Advanced forms of analytics and aritificlai intelligence are becoming increasingly deployed to support the work of healthcare workers. Medical doctors, nurses, and administrative staff either use, or are aided by sophisticated technologies which are posed to radically change the nature of their work. For example, radiologists now rely increasingly more on machine learning techniques to and other applications of AI to diagnose patients, while a lot of procedural and repeptive tasks are being done by machines. The objective of this project is to understand how the nature of work for health practitioners is changing, and what positive and negative consequences they experience.
This is a project in cooperation with Gintel.
This project aims to develop a pipeline for automated switchboard call handling using Speech-to-Text (STT), Text-to-Speech (TTS), and AI technologies. The context is automating inbound call traffic to a company. This typically includes:
This thesis uses Biometric data (heart rate, EEG, eye-tracking) to understand how the brain process visual conceptual models. Conceptual models are written in specific diagrammatic languages (two dimensional visual models) such as UML and BPMN. A lot of work has been done on the understanding of how humans comprehend and use such models in information systems and software development from the point of view of IT, cognitive psychology and linguistics. On the other hand, there is limited work on how the brain processes such models. Some work is done in neuro-lingustics, but primarily looking at natural language texts. Several areas of the IT-field has also used techniques from neuro-science for a while (NeuroIS where one look e.g. on the usage of IT systems and appropriate user interfaces, and NeuroSE where one in particular look on comprehension of software code)
The task consist in establishing an overview of current work in neuro-science, neuro-linguistics and IT relevant for understanding how the brain process conceptual models as part of model comprehension, and develop and conduct experiments to investigate aspects of this, including how visual conceptual models are processed under comprehension, if there are individual differences in the comprehension of models based on personal characteristics and if there are certain ways of modeling that are more appropriate than others seen from the point of view of human processing. This information can be used to develop neuro-adaptive modelling tools. The report is expected to be written in English, and a good master thesis will be able to contain material for a scientific publication.
Vi ansetter også forskningsassistenter inn mot tematikken . Oppgaven gjøres i samarbeid med Kshitij Sharma ved IDI og andre samarbeidspartnere ved NTNU og internasjonalt
This project asks a simple but important question: how can we computationally match brain structures across species - for example between mouse, macaque, and human - in a way that is not only geometrically plausible but also biologically useful?
Today, brain atlases and MRI templates are usually built separately for each species. That makes cross-species comparison difficult, even for brain systems that are highly relevant both in animal experiments and in human medicine. This is especially important for brain systems that are commonly sampled in humans with intracranial electrodes, such as parts of the extended basal ganglia and hippocampal systems, with Deep Brain Stimulation (DBS) and stereo-ElectroEncephaloGraphy (sEEG) electrodes, respectively.
The master thesis would focus on the computational side of this challenge. The student would work with 3D brain atlases, meshes, surfaces, labels, and landmarks, and compare different strategies for creating correspondences across species. Depending on the student profile, the work may involve geometric processing, image registration, machine learning, graph-based methods, quality-control metrics, visualization, or software engineering.
A realistic first goal is not to solve the full mouse-to-human whole-brain problem in one step, but to build and benchmark a reproducible pipeline that can be extended. This may include testing mapping methods on selected brain systems first, defining evaluation criteria, and identifying which technical approach is most promising for a larger research program.
The expected outcome is a well-documented benchmark pipeline, reusable code, visualizations, and a clear technical recommendation for future work. The project should be attractive to one or two MSc students in AI / computer science and would ideally be paired with a neuroscience MSc student in our group next year. Potentially, Adam Petro may also help bridge the neuroanatomy and computational sides as he is already doing an MSc in Neuroscience in this area and may continue toward a PhD.
The thesis will be carried out in collaboration between the Clinical Brain Systems (CBS) group at NTNU / St. Olavs Hospital and collaborators in computer science and AI at NTNU. CBS is a translational neuroscience environment working at the interface between clinical neurology, neuroanatomy, neuroimaging, and intracranial human brain recordings. The long-term goal of this project line is to build computational tools that make it easier to compare brain systems across species in a biologically meaningful way and to connect animal research more directly to human brain data.
External collaborator:
Name: Thanh Doan
Email: thanh.p.doan@ntnu.no
Internal supervisors: Gabriel Hanssen Kiss kiss@ntnu.no and Frank Lindseth frankl@ntnu.no
This project explores the feasibility of using AI tools to support the development of a small educational operating system inspired by xv6-riscv. The student will design and implement a simplified OS from scratch, covering selected core components such as process management, system calls, scheduling, memory management, and basic I/O. AI tools (e.g., coding assistants and large language models) will be actively used throughout the development process to support code generation, debugging, and understanding of system behavior.
A key aspect of the project is not only the implementation itself, but also the systematic documentation and evaluation of AI-assisted development. The student will analyse how AI contributes to different stages of OS construction, identify its strengths and limitations, and assess its impact on productivity and understanding of operating system concepts.
The project aims to serve as a preliminary experiment for educational innovation in TDT4186 Operating Systems. Based on the implementation experience, the student will propose how a similar AI-supported “build-your-own OS” approach could be integrated into the course, including suggestions for project structure, learning outcomes, and student guidance. The final outcome will include a working OS prototype, a documented development process, and a set of recommendations for using AI in operating systems education.
Over the past decades, there has been significant progress in digital accessibility, driven by better tools, stronger governance, and increased awareness. However, shifting economic priorities and limited understanding of accessibility concepts threaten to stall this progress. For many developers, accessibility remains a vague and complex area — a checklist of standards without a clear sense of how to meet them or why they matter.Accessibility is a broad field, encompassing diverse types of impairments — from visual and auditory impairments to motor limitations and cognitive challenges. Importantly, these impairments may be permanent, temporary (e.g., an injury), or situational (e.g., a noisy environment or glare on a screen). By designing with accessibility in mind, we improve digital experiences not just for those with disabilities, but for everyone.This project examines how interactive and educational games can be utilized to promote empathy and understanding of accessibility challenges. The idea is to simulate different impairments through playable web-based scenarios that highlight common accessibility failures. Players will experience the frustrations faced by users with impairments and then be guided through the process of improving the design, seeing firsthand how the same content becomes more usable and inclusive.Possible contributions of the project include:- Designing and implementing an accessibility-focused learning game- Evaluating learning outcomes or user experiences with a prototypeThis project is ideal for students interested in inclusive design, human-computer interaction, educational technology, or web development. It offers a chance to combine technical work with a meaningful social mission.
Co-supervisor: Dag Frode Solberg
Bakgrunn og motivasjon: For å møte klimautfordringer må Europeiske selskap integrere bærekraft i alle aspekter av forretningsdriften. Med innføringen av det nye Bærekraftsdirektivet fra EU (Corporate Sustainability Reporting Directive, CSRD), som trer i kraft i januar 2025, stilles det strengere krav til bedrifters rapportering av bærekraftsdata. Dette inkluderer detaljert rapportering av CO-utslipp, spesielt såkalte scope 3-utslipp, som omhandler indirekte utslipp i verdikjeden.
En viktig grunn til virksomheters utfordringer med bærekraftrapportering er at nøyaktig hva som skal dekkes av “bærekraft” er under-spesifisert, hvilke av virksomhetenes data som er relevant, hvordan data skal sammenstilles og hvem bærekraftrapportering er relevant for og hvordan.
Masteroppgaven utføres i samarbeid med Aneo AS (aneo.com, tidligere TrønderEnergi), et ledende nordisk fornybarselskap.
Målsetning: Dette masterprosjektet vil utforske utfordringer og løsninger knyttet til innsamling, behandling, og rapportering av bærekraftsdata for å møte de nye kravene.
Hovedveileder: Eric Monteiro, IDI/ NTNU
Medveileder: Kathrine Vestues, Aneo AS (kathrine.vestues@aneo.com)
Info om CSRD:
Problem Description: Despite the hype of AI development in recent years, true and impactful benefits of AI adoption in the public sector is still an unproven case. Reported adoption cases are often sporadic, random and lack the magnitude of impact at an infrastructure level technology change holds promise to. In this project, we set out to study the traits and characteristics of impactful AI-adoption in the public sector.
The thesis can involve:
OsloKommune: OsloKommune wants to get a 3rd party of up-and-coming students with state-of-the-art knowledge to analyse their newly established Oslo municipality’s AI factory. A factory consisting of modules that can be reused in new AI solutions. They want the students to familiarize themselves with the AI factory. Make an analysis, sketch and document the AI-factory, make recommendation on what can be done different. Students will work together with Oslo municipality’s centre of Excellence for Artificial Intelligence.
Value for Oslo Municipality:
Pool of Potential NTNU supervisors: Xiaomeng Su (technical), Casandra Grundstrom (sociotechnical)
We have several old systems that made it possible to ask natural language queries over Internet, by SMS or by voice over telephone about various tasks, e.g bus routes or telephone information. You can try it yourself at http://busstuc2.idi.ntnu.no
Telebuster is a prototype dialog system from year 2000, that understands conversations about travel information and can provide schedule information. The task will be to make it ready for 2026, by updating all the sub-systems to the latest technology, and by connecting it to GTFS, and recent large language models (LLMs) to make it available with updated schedules, anywhere in the world, all the time.
The existing system is built in layers with Prolog - PHP - and Java. It would be an advantage to reduce the pipeline by using only two layers. In light of the new LLMs, it would be interesting to explore how difficult/easy it is to get the 2000 dialog model up and running again (with the help of LLMs), vs implementing the same (or better) functionality with “modern approaches”. In terms of Sustainability, it should also be documented how much power is needed to run the old system vs the new systems.
Fluent knowledge of Norwegian and English is a huge advantage, but clever use of Google translate etc. can compensate for lack in experience.
Description:A web+mobile based system designed to help restaurants and food businesses manage internal control requirements for food safety (IK Mat) and alcohol regulations (IK Alkohol). The platform supports documentation, daily routines, and compliance with Norwegian laws, enhanced with AI-assisted tools.
Key Features:
Note - Collaboration with potential industrial customers is possible for the right candidates.
Context: RISC-V is an open-source instruction set architecture (ISA) offering a customizable framework for designing processors. RISC-V promotes the capability to define custom ISA extensions to customize processors for specific applications or performance requirements.Problem: Silicon is fixed in atoms, and so are compiler targets. How can compilers help a designer who is still in the process of defining the capabilities of a new architecture in terms of ISA features? What if we want to explore a set of alternative designs?Goal: We want to make the LLVM compiler a valuable tool in the early stages of hardware design by being able to dynamically define custom compilation targets. For example, if a designer must create the best accelerator for a specific domain (e.g., autonomous driving or post-quantum cryptography), what are the most useful vector instructions to accelerate? Can the compiler recompile the same program for a wide range of variations of the same ISA and assess the impact on performance?Requirements:Programming Languages: C/C++ (mandatory), Python (desired)Tools: CMake, GitOS: *nixCompiler toolchain: LLVM (desired)English language: working proficiency (mandatory)
This project involves regular collaboration with:
Classifying animals in the wild in complex backgrounds is a challenging and open problem. The complexity increases with natural vegetation, varying environmental conditions, half-captured animal parts in frames, lightning variations, and so on. This project aims to classify not only the animals captured in the frame (i.e., elk, rabbit, deer) but also the background surroundings into snow, grass, trees, etc. The project entails developing deep learning models to label foreground objects (animals) and the background into distinctive categories. Elephant Expedition (EE), Snapshot Wisconsin (SW), Snapshot Serengeti (SS), and Camera catalog (CC) to identify animal species in the wild are some of the datasets that contain millions of images that can be used for training and testing deep learning models. Norwegian wildlife dataset collected by NINA is also available with us.
With limited knowledge of an ISA, and many examples of firmware binaries, how can code re-use and binary diffing be effectively used as a tool to help the reverse engineer?
This project aims to take publicly available information such as multiple versions of firmware binaries and scant changelogs to deduce instruction set architectural (ISA) information about proprietary hardware platforms. We continue some initial work on a case study involving Shimano STePS electric bicycle firmware, which runs on proprietary embedded hardware.
Research questions include:
Related work: https://dl.acm.org/doi/10.1145/3555858.3555948
Approximate computing is the science that studies how to provide ‘good enough’ results -- according to an application-specific quality metric -- while at the same time improving a performance metric such as time-to-solution, energy-to-solution, area, etc.
Among many approximate computing techniques, precision tuning focus on modifying the data types used in the computation in order to reduce the number of bits used to represent real numbers, and/or replace floating point data types with fixed point numbers or other numeric standards, depending on the hardware capabilities.
Precision tuning requires a various range code analyses, transformations, and verification tasks, each of which can be subject of small or big project ideas.
This project will involve international collaborations with:
Main contact: Stefano Cherubin
Equation-based modelling and simulations languages offer a high-level interface to allow modellers to define the behaviour of a dynamic physical system in terms of Differential Algebraic Equations (DAEs), a key element in the simulation and digital twin approach to system modelling.
Numeric integration algorithms can be used to obtain a description of the evolution of the physical system over time by observing its key elements (state variables).
The current state of the art allows for automatic translation of declarative models into executable simulations. However, the current translation algorithms do not scale with large number of state variables.
Ongoing research efforts aims at:
Other contacts: Tor Andre Haugdahl
™Computational thinking involves solving problems, designing systems, and understanding human
behavior, by drawing on the concepts fundamental to computer science.” (Wing 2006, https://www.cs.cmu.edu/~15110-s13/Wing06-ct.pdf). Computational thinking is now recognised as a basic skill that should be developed since early education.
This task aims at exploring the evolution of computational thinking as the use of AI tools in schools increases. How is this concept changing? The specialization project will focus on building the core background knowledge through a literature reviews and, possibly, interviews with relevant stakeholders.
The project can then be continued in the master by designing and evaluating a (computer-based) intervention to develop computational thinking with the help of AI tools, promoting ethical and critical reflection.
The task is particularly relevant for students of the program Lektorutdanningen i realfag or to students in other programs with an interest in technology in educational settings.
The task will be co-supervised with Erica Perseghin, University of Udine, Italy
Credit risk modeling is critical for financial institutions managing unsecured credit portfolios. This research focuses on modeling the three primary components of credit risk: Probability of Default (PD), Loss Given Default (LGD), and Exposure at Default (EAD) for a bank with over 1 million credit card customers. The study will incorporate both micro-level (customer-specific) and macro-level (economic) factors affecting default risk. This proposal outlines the research objectives, data sources, methodology, and expected contributions. Research Objectives:
- Develop robust statistical and machine learning models to estimate PD, LGD, and EAD for credit card customers.
- Analyze the impact of micro-level characteristics (age, sex, employment status, transaction history, etc.) on individual default probability.
- Examine macroeconomic influences (GDP, inflation, interest rates, unemployment) on aggregated default rates.
- Leverage unique internal bank transaction data and Gjeldsregisteret data to improve model accuracy.
- Assess the impact of credit card reward programs on default behavior.
- Provide policy recommendations to financial institutions for better risk management.
Existing literature on credit risk modeling primarily focuses on PD estimation, while LGD modeling is relatively underexplored. Studies suggest logistic regression as a baseline for PD modeling, while advanced machine learning techniques (e.g., Random Forests, XGBoost, Neural Networks) have shown promising improvements. The Norwegian market presents unique challenges due to historically high debt recovery rates, making LGD modeling particularly important. Additionally, the role of credit card reward programs in influencing risk behavior remains an open research question.
Geographical Information Systems (GIS), such as ARCGIS, provide a platform for crowdsourcing information from a wide geographical area. This approach has been used to crowdsource geological and climate related content, as well as narratives about the specific locations. This project aims to create a crowdsourcing GIS platform that could contribute to enhancing the knowledge about places through sharing stories and interesting experiences that would showcase a place and contribute to providing a sense of a place. Furthermore, the use of Large Language Models (LLM) should be explored. The sub-tasks include:
Data-driven data science is attracting a lot of interest.
However, uptake into organizational practice is lagging significantly behind. Why?
The focus of this project/master is to supplment the possibilities provided by data-driven data science techniques with an empirical understanding of the conditions and circumstances for these techniques to be used in practice for consequential decision-making.
Empirical cases to study data science in practice we will have to discuss. Candidates include: the energy sector, healthcare
An example problem situation is the efforts to enhance the accountability and explainability of algoritms (XAI, explainable AI).
The project/master is part of SFI NorwAI (https://www.ntnu.edu/norwai) funded by the Norwegian Research council
Økt utenforskap blandt unge belyses som et økende samfunnsproblem av både forskere og offentlige ansatte. Funn fra forskning tyder også på at unge har manglende forståelse om hvordan velferdstjenester er organisert, noe som hinderer de å få hjelpen de behøver, og jo lengre man står utenfor, desto større utfordringer møter man i forsøk på å komme i jobb.
I dette prosjektet ønskes det å jobbe direkte med unge (mellom 16-25) for å utvikle digitale verktøy og metoder som kan gi unge en reell stemme og styrket rolle i utformingen av velferds- og helsetjenester. Dette vil ikke bare kunne gi unge bedre støtte og økt myndighet mens de navigerer det offentlige hjelpeapparatet, men også kunne bidra til kunnskap om hvordan tilbudet av velferdstjenster kan bedre tilpasses og potensielt lette overgangen for unge til et aktivt liv. Samtidig er det viktig at personvern sikres.
Det er en fordel om du har interesse og kunnskap om empirisk, kvalitativ forskning og designarbeid. Oppgaven kan leveres på norsk eller engelsk, men god norsk beherskelse er nødvendig.
Kontakt Tangni Dahl-Jørgensen, tangni.c.dahl-jorgensen@ntnu.no, for mer informasjon om prosjektet.
How can we help informatics students to get a better understanding of the impact of the technology they develop? This task will focus on designing a playful approach for learning about sustainability of IT solutions and how to integrate sustainability awareness in IT design.
The task will start from the co-design toolkit Tiles (https://www.tilestoolkit.io/) to modify it to promote reflection on sustainability. Students might choose to develop a physical, online, or hybrid toolkit.
Previous work has been done in the group about teaching about sustainability to informatics students and provides a good starting point, still giving freedom to shape your work.
The task is expected to include design, prototyping and evaluation.
Contact the supervisor to share your ideas and know more about this task
In today's digital age, it's crucial to manage important documents like diplomas and licenses securely and efficiently. Traditional methods of handling these documents are either outdated i.e. paper based or fragmented and can pose privacy and security risks. However, with new web 3.0 technology like self-sovereign identity and digital wallets, there's an opportunity to improve how we manage identity documents. This proposal aims to introduce a digital wallet platform that can securely store various identity documents such as academic diplomas, driving licenses, boat licenses, flying licenses, shooting licenses, and so on. By using advanced technology, this platform will make it easier for users to access and manage their documents while ensuring their privacy and security.
The main goal is to create a digital wallet platform (mobile + web) that can safely store and manage a wide range of identity documents. This platform will serve as a centralized place for users to keep their important identities documents, reducing the need for physical copies and/or fragmented storage methods. Additionally, the platform will include additional layers of security features, like encryption and biometric authentication, to protect users' sensitive information.
Furthermore, the envisioned platform will be user-owned, user-friendly, with easy navigation and integration with other digital systems. Users will be able to upload, organize, and share their documents with relevant authorities quickly and securely. The platform will also provide automatic reminders for document renewals, helping users stay compliant with regulations.
Here's a summary of the proposed tasks:
Summary:This project aims to design and test the feasibility of an inclusive mobile application platform to support the mental health of mothers caring for children with intellectual disabilities. The application platform will offer AI powered culturally adapted, low-literacy-friendly tools, including visual resources, local-language content, stress management support, and private connections to therapists - tackling barriers of stigma, access, and cost.
Activities:
Conduct needs assessments through focus groups and surveys
Develop a basic prototype with AI-driven stress management and therapist directories
Test usability, acceptability, and technical performance
Collaborate with experts in health innovation and mobile app development
Collaborations:
Namrata Pradhan (namrata.pradhan@ntnu.no) from the Department of Mental Health will serve as the product owner, providing support for refining project requirements.Surya Kathayat from the Department of Computer Science will act as the project supervisor.
The urgent need to reduce carbon emissions from maritime activities has highlighted the importance of innovative strategies in interaction design, particularly eco-feedback, which has been effective in nudging car users towards more efficient practices. Despite the acknowledged potential for significant emission reductions on ships through behavioral changes in their operation, there are currently no established standards for eco-feedback within the maritime sector. The project will focus on the development of new concepts and ideas for providing eco-feedback interfaces to ship operators, aiming to encourage a shift towards more sustainable behaviors.
Key activities in this project will include:
This project is part of the OpenZero project, which aims to reduce carbon emissions in maritime sector. The student will receive support from the partners affiliated with the project.
The project will be co-supervised by Dr. Taufik Akbar Sitompul (Department of Design, NTNU).
Research assistant job possibility: There is the possibility to enroll as Research Assistant for tasks related to the context of this project. Please get in contact if you want to discuss this opportunity: leonardo.montecchi@ntnu.no
The rise of Open Science (OS) represents a profound and impactful change in how research is conducted and disseminated. According to the UNESCO Recommendation, open science is defined as a set of principles and practices designed to make scientific research from all fields accessible to everyone [1]. OS enhances research transparency, and acceleratesing scientific progress through openness, collaboration, and innovation.
The integration of Artificial Intelligence (AI) within this movement further expands its potential, enabling more efficient knowledge discovery, enhanced reproducibility, and deeper interdisciplinary research. However, realizing this vision requires skill development for researchers and other actors with Higher-Education Institutions (HEI). Currently, inconsistent adoption of Open Science policies across institutions hinders international collaboration and research exchange [2].
The AIOS Erasmus+ Project looks at the interactions between Open Science (OS) and Artificial Intelligence (AI), with the objective to define a curriculum for OS with and within AI. Within this context, this master project focuses on developing and evaluating two self-assessment tools to support the teaching of skills to integrate AI and OS.
Self-Assessment Tool: This tool is designed to empower learners by providing them with the means to assess their current knowledge, skills, and attitudes relevant to open science practices. It supports learners in measuring how well the content aligns with their learning objectives, applying new knowledge immediately, testing their understanding of new concepts, and receiving constructive feedback.
Learning Journeys Design Tool: Complementing the self-assessment tool, this innovative tool facilitates the creation of customized learning journeys. These journeys are tailored to the specific competency levels identified through self-assessment. This personalized approach ensures that the learning experience is as effective and impactful as in-person education while maintaining the flexibility and accessibility of online learning.
The work in this project will be in connection with partners from the AIOS project, which involves institutions in Greece, Italy, and Ukraine, besides NTNU.
Web3 technologies open new and interesting possibilities in games development! In this project a gaming platform will be developed that allows multiple players to play, learn and/or earn points or digital assets.
Playing activities can be different depending on the chosen game (Student shall propose a game!!) For example
The platform will also encourage users to create games or game contents and reward them!
Mechanisms for exchanging rewarded points with other digital assets shall also be proposed and implemented!
Possible research aspects:
The way we write texts give a lot of information about the background personalities of the authors: their age, gender, native language (if writing in a foreign language), if they're human or bots, and possibly their actual identity. This type of information can be used to, e.g., give fair indications of user profiles, to deduce if a text (or a part of it) has been plagiarised, or to uncover social media software misuse. The thesis could thus focus on tasks such as author profiling (what can we say about the author, e.g., their gender, age, if they're a human or a bot), author identification (did a specific author write this text?) and/or plagiarism detection (did somebody else than the author claiming the text actually write all or part of the text?), looking at textual data from, e.g., social media sites, chat rooms or parliamentary debates, and apply machine learners such as Transformer technologies and Large Language Models to the texts in order to draw conclusions about who the author behind a text is.
Read also: Writing a Master's Thesis in Language Technology
Nyankomne til Norge må gjennom mange byråkratiske søknadsprosesser som krever en høy grad av systemforståelse. Samtidig er dette er en diversifisert brukergruppe som har ulike utgangspunkt mtp alder, utdanningsnivå, og språklige og digitale ferdigheter, noe som kompliserer tilgang og forståelse til informasjon som er nødvendig for å få varig oppholdstillatelse og et godt liv i Norge.
Dette mastergradsprosjektet går ut på å samle brukerperspektiver om flyktningers og immigranters utfordringer i innhenting av informasjon om eller bruk av offentlige tjenester, og utføre deltakende design/codesign aktiviteter i utviklingen av digitale støtte, f.eks. i form av samling av ressurser, forenklet informasjon innhenting, gamification for barn og unge etc. Her er det også mulighet for å samarbeide med offentlige, kommunale tjenestetilbydere og/eller frivillig organisasjoner som jobber med flyktninger til daglig.
Oppgaven kan leveres på norsk eller engelsk, men det kan være en fordel med grunnleggende norskforståelse.
Prosjektet krever at du er interessert i menneskelig aspekter ved utvikling av nye teknologiske løsninger og kan tenke deg å jobbe sammen med mennesker i en vanskelig livssituasjon på en måte som er myndiggjørende og ivaretar brukerens medvirkning i design og utvikling.
Reverse engineering (RE) is the process of discovering features and functionality of a hardware or software system. RE of software is applied where the original source code for a program is missing, proprietary, or otherwise unavailable. Motivation for RE ranges from extending support of legacy software to discovery of security vulnerabilities to creating open source alternatives to proprietary software.
RE usually targets binary programs with a known instruction set architecture (ISA) and executable format. The RE process proceeds by disassembling the binary into assembly code, and where possible decompiling the assembly to yield high-level source code (for example, C source code).
However, in many cases the ISA is either undocumented, unknown, or unavailable. In addition, malware has been shown to use custom virtual machines to avoid detection. Such cases prove extremely time intensive for the reverse engineer. ISA features such as word size, instruction format, register size, and number of physical registers are a prerequisite to disassembly.
The task for this project is to develop a heuristic method to analyse binary programs and extract instruction set architectural features such as:
Knowledge of computer architecture and assembly programming is helpful but the lack of such knowledge should not discourage the applicant(s). Strong Python programming skills are desirable.
The task for this project is to develop a heuristic method to analyse binary programs and extract subroutine structure (e.g., CALL/RET instructions), branch and jump statements, entry points, etc. to recover program control flow.
Previous work: https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3120623
As LLMs are increasingly becoming a commodity - “everyone” are making them - a big, remaining challenge is to use LLMs for specific tasks, in a particular domain.
This project looks at the work we are doing at NorwAI to pre-train and fine-tune Norwegian LLMs for designated use-cases in healthcare in Helse Midt-Norge
After the introduction of ChatGPT and other generative AI models (LLMs), code generation using AI has become more and more popular. While initially simple textual prompts were used, it was soon clear that those would provide sub-optimal performance and would be more prone to causing “hallucinantions” in artifacts generated by AI models [1].
Both research and the industry are moving to more structured prompts, with processes where prompt instructions are provided as structured data in Markdown or JSON format, or with frameworks such as DSPy [2] or PDL [3].
Domain-Specific Languages (DSLs) [4][5] are tailored programming or modeling languages designed for a particular application domain, offering higher-level abstractions than general-purpose languages (GPLs). In classical software development, Domain-Specific Languages (DSLs) have been used as intermediate format between high-level textual requirements and lower-level code in traditional programming languages. They have been shown to improve productivity, reduce errors, and enable domain experts to express solutions in a more natural way.
The objective of this project is to investigate to what extent using DSL in combination with Generative AI can improve the accuracy of code generation tasks.
The project aims to revolutionise university credit management by integrating blockchain with existing educational systems using Blackboard and Inspera APIs. EduWallet ensures secure, efficient record-keeping and easy credit transferability across institutions.
API Integration:
Smart Contracts for Enrollment: Simplifies the course enrollment and withdrawal processes using blockchain to guarantee secure, verifiable transactions.
Credit Transfer: Employs a blockchain ledger for secure storage and seamless transfer of credits, enhancing student mobility.
Real-Time Verification: Offers instant verification of academic records, accessible by employers and institutions, ensuring accuracy and preventing fraud.
Token-Based Incentives: Rewards students with tokens for academic achievements, redeemable for special privileges like exclusive workshops or priority course enrollment.
Analytics Dashboard: Provides a real-time dashboard for students to monitor credits, course statuses, and rewards eligibility.
Web app (and or Mobile app), Blockchain, API integrations, Database, Security
Let's make a [MLIR]-based compiler for visual block-based programming languages (e.g., Scratch, Blockly).
The target architecture is a Parallel Ultra-Low-Power RISC-V architecture developed by University of Bologna & ETH Zurich [architecture].
This project is about making a compiler. To start this project you need strong C++ and low-level programming skills.
This project is a follow-up from a previous year project and you are expected to continue the development of an existing toolchain.
Requirements:
[MLIR] https://mlir.llvm.org
[architecture] https://arxiv.org/abs/2301.03904
Vision-Language Models (VLMs) combine visual understanding and language reasoning, enabling tasks such as visual question answering, scene understanding, object grounding, image captioning, and human-robot interaction. These capabilities make VLMs highly attractive for edge AI applications, especially in systems that must interpret visual environments and respond intelligently in real time.
However, VLMs are typically large and computationally demanding. Their deployment on embedded platforms such as NVIDIA Jetson AGX Orin is challenging due to limited memory capacity, power constraints, and latency requirements. In particular, the combination of a vision encoder and a large language component often results in heavy computation and high memory traffic, making naive deployment impractical.
This project investigates how to optimize VLMs for edge deployment on NVIDIA AGX Orin. The goal is to reduce latency, memory footprint, and energy consumption while preserving task performance as much as possible.
Deploying VLMs on NVIDIA AGX Orin introduces several key challenges:
Unlike conventional vision-only models, VLMs involve interactions between visual tokens and language tokens, which makes optimization more complex. Efficient deployment therefore requires model-aware and system-aware strategies.
The main objectives of this thesis are:
1. Literature Review
The student will first review:
2. Baseline Deployment
A baseline VLM will be selected and deployed on NVIDIA AGX Orin using frameworks such as PyTorch, ONNX, and TensorRT where applicable.
Typical tasks may include:
The baseline system will be profiled in terms of:
3. Optimization Technique
The project will investigate a subset of the following techniques:
Apply FP16 or INT8 quantization to reduce memory usage and improve inference speed
Reduce redundant parameters or layers in the visual encoder, projector, or language decoder
Reduce the number of visual tokens or intermediate representations processed by the model
Improve autoregressive decoding efficiency by reducing unnecessary language-side computation
Analyze whether different components can be scheduled efficiently across available hardware resources
Compare alternative compact VLMs or smaller vision/language encoders suitable for edge deployment
The student is not expected to implement all techniques. The thesis should focus on one or two optimization directions in depth.
4. Benchmarking and Evaluation
The optimized system will be compared against the baseline using metrics such as:
The evaluation should highlight the trade-offs between efficiency and model capability.
5. Deployment Study
The final part of the project will examine how well the optimized VLM can support real-time edge applications on NVIDIA AGX Orin. This may include a simple demonstration scenario, such as:
Expected Background
A student taking this project should ideally have:
The thesis project would explore either emotion recognition or automated generation of artworks (music, images, videos or texts) tailored to some emotion - or a combination of those two themes. In either case the work needs to explore (general) emotion taxonomies, data (music, art, poetry, etc.) with and without emotional annotations, and techniques for emotion classification in the chosen artform(s), tentatively utilising Transformer-based (Large Language Model) technology for the task of analysing and/or generating the texts, images or music scores.
Read also: Writing a Master's Thesis in Computational Creativity
This thesis is part of INCLUDE‑ID, a Horizon Europe research project dedicated to transforming cultural heritage institutions into inclusive, adaptive environments for children with intellectual disabilities (ID).
Supervisors: Michail Giannakos, Giulia Cosentino
IntroductionCultural heritage and schools are increasingly integrating digital solutions to make their environments more accessible, yet children with intellectual disabilities (ID) remain among the most underrepresented groups. Their participation is often limited by cognitive, sensory, communicative, and environmental barriers. The INCLUDE‑ID project aims to address this challenge by exploring how adaptive and multisensory technologies can personalize cultural experiences based on each child's functional profile. By leveraging AI, multimodal interaction, and data‑driven personalization, cultural spaces can become more inclusive, engaging, and autonomy‑supportive for visitors with ID.
This thesis investigates how digital adaptive technologies can personalize cultural experiences for children with intellectual disabilities within cultural heritage institutions. Building on the INCLUDE‑ID framework, the student will explore how AI‑driven personalization, multimodal interfaces, and multisensory design can dynamically match the characteristics of a cultural space with the needs, preferences, and abilities of each user. The work involves reviewing research on accessibility technologies, adaptive systems, multimodal learning analytics, and inclusive design; developing a conceptual or computational model that connects user profiles with environmental features; and designing a prototype or simulation that personalizes content, navigation, or interaction modes. The aim is to identify technical mechanisms that support autonomy, participation, and meaningful engagement, contributing to replicable adaptive design practices for inclusive cultural heritage environments.
RequirementsThe ideal candidate has experience in AI, machine learning, or multimodal interaction, and an interest in accessibility and inclusive design. Strong programming skills (Python, C#, or JavaScript) and familiarity with data processing, interaction design, or prototype development are beneficial. Motivation for hands‑on experimentation and interdisciplinary collaboration is essential.
This Master’s project focuses on the design and implementation of an end-to-end digital infrastructure for collecting, storing, and visualizing health-related data from wearable devices and Patient Reported Outcome Measures (PROMs). The system will support monitoring and evaluation of care outcomes in scenarios where multiple caregivers—such as clinicians from different specialities, therapists, and family members—are involved in the treatment process.
The project addresses key challenges in modern digital health systems, including heterogeneous data integration, real-time data processing, interoperability, and user-centered visualization. Wearable devices generate continuous streams of physiological and behavioral data (e.g., activity levels, heart rate, sleep patterns), while self-reported data provides subjective insights into symptoms, well-being, and treatment adherence. Combining these data sources into a unified platform enables a more holistic understanding of patient outcomes providing the patient with an overview over the outcomes of different treatments and, allows for a learning mechanism for health care professionals.
A central aspect of the thesis is to design an architecture that ensures reliable data ingestion, secure storage, and scalable processing pipelines, while enabling meaningful visualization tailored to different stakeholders. Special attention will be given to supporting multiple caregivers with varying information needs, ensuring that the system presents relevant insights in an accessible and actionable manner.
The thesis will also consider aspects such as data privacy, access control, and ethical handling of sensitive health information, as well as system usability and Quality of Experience (QoE) for both patients and caregivers.
The thesis may result in:
Objective: To develop and evaluate AI-based analytics models for tracking and enhancing learner engagement and performance in simulations, focusing on STEM.
Description: This research aims to develop AI-based analytics models to enhance the effectiveness of laboratory simulations in online learning. The study will focus on tracking learner interactions within simulations developed using Articulate Storyline 360, generating actionable insights into student engagement, skill acquisition, and performance trends. These insights will inform real-time interventions and instructional design strategies to improve student outcomes. A case study on STEM education.
Smart and personalized systems such as recommender systems (or artificial intelligence in general) keep influencing our daily lives in an increasing rate. In the recent years, researchers became more aware of the ethical challenges in developing such systems in an ethical way such that these systems would treat everyone equal and fair, without any bias or discrimination. Even though these are the topics social scientists have been working on for a long time, defining these concepts as mathematical models, implementing them within AI systems and evaluating the success of these approaches is not an easy task.
This project focuses on the ethical aspects such as privacy, fairness and bias in recommender systems from the technical point of view. The direction and details of the project can be clarified upon a discussion with the students.
A few references on the topic:
https://blog.fiddler.ai/2021/03/ai-explained-understanding-bias-and-fairness-in-ai-systems/
https://www.ibm.com/blogs/research/2021/04/ibm-reduce-bias-in-healthcare-ai/
AI, Inclusion, and Volunteer Practice: An Ethnographic Study of Student Engagement in AINCLUSION
This master's thesis invites motivated Computer Science students at NTNU to combine volunteer work, AI expertise, and societal engagement by contributing to the activities of AINCLUSION (https://ainclusion.com). AINCLUSION is a non-profit association that promotes inclusive, accessible, and responsible approaches to artificial intelligence in society.
The student will participate as a volunteer in AINCLUSION's ongoing activities (workshops, events, community outreach, digital content, or other initiatives). Through this engagement, the student will conduct an ethnographic study focusing on:
Ethnography will serve both as a methodological framework and as a reflexive practice. The student will maintain field notes, document interactions, collect feedback from participants, and analyze how volunteer work contributes to the association’s goals of inclusion, empowerment, and AI awareness.
This thesis is particularly suitable for students who:
Strong foundational knowledge in artificial intelligence is required in order to analyze AI-related content, evaluate the technological aspects of the association’s activities, and interpret how AI impacts the communities AIinclusion engages with.
The thesis will use ethnographic methods, such as:
Ethical considerations—especially regarding consent, privacy, and the handling of qualitative data—must be documented and addressed in line with NTNU guidelines.
The thesis will be evaluated based on NTNU’s standard criteria, such as:
The student will work closely with:
Objective: To create gamified simulations with adaptive AI mechanics in Articulate Storyline 360 and evaluate their effectiveness in enhancing motivation and promoting deeper learning.
Description: This topic investigates the use of gamification in simulation-based learning environments developed with Articulate Storyline 360. The study will incorporate AI elements such as performance tracking and adaptive game mechanics to personalize learning experiences. By examining learner motivation, engagement, and retention rates, the research will assess how gamified simulations can address challenges in online education and promote deeper learning.
Are you interested in working with state-of-the-art AI accelerators? If yes, this project is for you. The project is about the evaluation of AI hardware accelerators. There are several open-source hardware implementations, but there is no unified framework that enables the developer to test and evaluate these accelerators.
The starting point of this project is the evaluation of Gemmini, an accelerator from Berkeley Architecture Research Lab. Gemmini is part of Chipyard/FireSIM, a framework that provides an easily extendable infrastructure for hardware designs. After evaluating Gemmini, the plan is to choose and evaluate another accelerator. There is a wide range of accelerators that can be integrated into Chipyard:
More specifically, the goal of the project is the following:
Further Reading
Prerequisites: TDT4295, TDT4255, TDT4260, or similar
Are you interested in working with state-of-the-art fast machine learning inference? If yes, this project is for you. A wide range of optimization and transformation passes can be applied to the AI inference graph. Many of these are essential for making machine learning feasible, given its heavy computational and memory demands. For example, transformations such as ImageToColumn enable convolutional and transformer neural networks to run efficiently on hardware accelerators that only support matrix multiplication. In addition, there is a broad set of optimizations tailored to extremely low-bitwidth arithmetic (below 4-bit).
Starting from an ONNX graph [5] and compiling down to C code is possible with several open-source tools [1,2]. But the very efficient and fast inference happens at low bit-width. QONNX [3] extends the ONNX graph with relevant operations/nodes. Compiling down to C code from QONNX to execute efficient memory and inference in microcontrollers is the first part. The second part is to apply the correct sequence of transformation passes to make the computation efficient. This approach can be extended beyond microcontrollers to a wide range of accelerators, from systolic array designs to dataflow accelerator architectures[4].
References (Further reading)
Prerequisites: TDT4205, TDT4295, TDT4255, TDT4260, or similar
AI systems are increasingly applied to inform decisions in central government agencies. If these decisions directly impact natural persons, they have to be explainable by law. The thesis should investigate to what extent non-generative black box AI systems can be used in decision support systems in the Norwegian central government. Possible research problems:
The task is done in collaboration with The National Audit Office of Norway (Riksrevisjonen) AI systems are increasingly applied to inform decisions in central government agencies. If these decisions directly impact natural persons, they have to be explainable by law. The thesis should investigate to what extent non-generative black box AI systems can be used in decision support systems in the Norwegian central government. Possible research problems:
The task is done in collaboration with The National Audit Office of Norway (Riksrevisjonen). The National Audit Office of Norway is the government’s auditing and monitoring body, contributing to the monitoring of democracy to ensure sound and effective management of the state’s resources in accordance with the Storting’s decisions and prerequisites.
This task is done spring 2026, extension for 2027 is not decided
AI governance is a notion that is often attributed to a range of different practices and processes. From establishing a process of developing AI applications, ensuring that quality outcomes are achieved, and to decising the role and responsibilities of stakeholders. AI governance now plays an important part related to the business value that AI can deliver, and to ensuring that projects comply with ethical and regulatory frameworks. This project will seek to understand how organizations develop AI governance practices, what aspects they take intio accoutn when doing so, how they deploy them, and what the outcomes of them are at the business and project performance levels.
The impact of news articles on the society can not be underestimated and as the number of online news are increasing, distinguishing the fake news from real news is becoming a challenge for people. This project focuses on analyzing and/or tracking news articles from different news sources or social media channels, in order to find an efficient way of detecting fake news.
In this project different methods (such as machine learning, natural language processing, semantic web etc.) or approaches can be used towards the detection of fake news / disinformation.
The details of the project can be clarified upon a discussion with the students.
Figurative language is used when the intended meaning of a statement isn't necessarily the one shown on the surface, that is, when the language intentionally conveys secondary or extended meanings, such as sarcasm, irony and metaphor. Such intentional ambiguity is also a key part of many jokes. Understanding and generating figurative language create significant challenges for language models, as direct approaches based on words and their lexical semantics often are inadequate in the face of indirect meanings. The project could thus focus on one specific type of figurative language (e.g, sarcasm or humour), and either investigate models that could interpret such figurative language or that could generate it.
Formålet med prosjektet er å forsøke å forbedre transkripsjonsnøyaktigheten til NB-Whisper, en tale-til-tekst-modell utgitt av Nasjonalbiblioteket, for inngangssamtaler i Sykehjemsetatens helsehus. For å oppnå dette gjennomføres fine-tuning av modellen med data bestående av lydopptak og tilhørende transkripsjoner fra slike samtaler. Samarbeidspartner i prosjektet er Helseetaten i Oslo Kommune der det er viktig å undersøker hvordan ASR-feil kan påvirke klinisk tolkning i en end-to-end klinisk transkripsjonspipeline, der nedstrøms NLP (Natural Language Processing)-komponenter (kan for eksempel være chatGPT) avhenger av nøyaktigheten til ASR-utdataene. Virkningen avhenger av typen feil snarere enn den totale feilraten. Det kan være interessant å se hele "prosessen".
Multiple movements like opening or closing the hand, grasping, or showing the palm can be decoded from the EEG signals recorded while attempting to do those movements. The decodified movements can serve for multiple purposes. For example, in neurorehabilitation they can be used to provide feedback to a patient that is performing therapy to recover hand movements after stroke, and in brain-computer-interfaces to generate outputs that control an external device such as home appliances, computer games, toys.
The objective of this project consists of decoding movement intentions by combining low-density EEG and source reconstruction (estimation of the activity inside the brain from the electrodes on the scalp). The project involves recording EEG signals for multiple participants, analyze and build offline/online classifiers using state-of-the-art machine/deep learning algorithms and developing software games such as the one in this link: Video semi final (youtube.com).
Preliminary results are now published here: https://link.springer.com/article/10.1186/s40708-024-00224-z
This project will provide a foundation to develop wearable solution based on few electrodes that can be later use in neurorehabilitation therapies. The project is done in collaboration with Marta Molinas at the cybernetics department
Visuospatial neglect are commonly experienced neuropsychologicalconditions affecting the contralesional side in post-stroke patients, leavingpatients with egocentric or allocentric perceptual problems. Diagnostic tools forvisual neglect include the apples test, balloons test, and bells cancellation test.All administered on paper. While psychometrically sound, these tests areadministered in an overtly clinical setting, lacking depth as a test parameter,only allowing for crude temporal data collection and gaze observation, andalso being limited in spatial scope to the size of the paper. By having a limitedset of test parameters, the status quo represents a barrier to advancing ourunderstanding of the mechanisms behind visouspatial neglect and its effect ineveryday settings of the patients. To mitigate these limitations, a VRenvironment is being developed for assessment of neglect by utilizing low-cost,off-the-shelf, VR headsets with integrated eye trackers, along with a custom-developed and highly flexible virtual environment, where the test parameterscan be altered based on the needs of the clinician.This master project aims to bring flexible data visualization into theaforementioned VR environment. For starters, the students will look at usingPython to generate graphs/plots of gaze data which has previously beencollected from the VR environment, and then to display these plots directly inthe VR environment itself (i.e. the gaze plot of the person during the test).Next, one can classify the fine-grained gaze data into more meaningful unit,such as fixation or saccade, and come up with more relevant staticalinformation to the clinician. Eventually the information should be visualized inan easy to comprehend manner to the clinician.This work collaborates with Department of Acquired Brain Injury, St. Olav'sHospital.
Supervisor: Alexander Holt
Co-supervisors: Tor Ivar Hansen, Xiaomeng Su
This project is part of a joint initiative funded by the Peder Sather Grant at UC Berkeley. It will be co‑supervised by researchers from UC Berkeley and Dr. Per Håkon Meland at SINTEF.
Problem statement
Generative AI (GAI) technologies, particularly LLMs, are known to have weaknesses, including hallucinations, misinformation, and disinformation. LLMs are also unsafe because they are vulnerable to various attacks and abuses. The recent development of integrating AI as agents and components into a system of systems has led to more unpredictable system behavior due to indeterministic, unexplainable LLM outputs. In particular, AI technologies could be abused to enable novel cyberattacks, leveraging their efficiency and intelligence.
Analyzing AI abuse risks beyond the foundation models themselves is essential, as many cases of abuse occur in software systems that integrate the models. Additionally, it is critical to be proactive in innovating and applying automated approaches and tools to identify and mitigate the AI abuse risks to defend against the negative consequences of compromises.
Task description
This master's thesis project is expected to contribute to the overall project objectives, including:
1) A systematization of knowledge of abuse of emergent GAI technologies;
2) An enhanced AI Risk-Management Standards Profile for General-Purpose AI (GPAI), Foundation Models, and Systems integrating the Models;
3) A set of approaches and tools to automatically identify and mitigate abusive risks of GAI technologies.
The pre‑project may focus on summarizing the state of the art to identify specific abusive risks associated with GAI technologies. The master’s thesis can then propose and pilot approaches to defend against these risks.
Basic knowledge of AI and GAI technologies.
In an era defined by rapid technological change, global challenges, and evolving labor markets, entrepreneurship and innovation skills have become fundamental competencies for graduates across all disciplines. These skills, which include opportunity recognition, creative problem-solving, risk management, resilience, and adaptability, are no longer solely the domain of business founders but are increasingly sought after by employers in all sectors. However, a significant gap exists between this growing demand and the current state of higher education. Traditional pedagogical approaches often struggle to convey the iterative, uncertain, and action-oriented nature of entrepreneurial thinking. Students need experiential learning environments where they can safely experiment, confront failure, and develop the mindsets required to navigate complex, ambiguous problems. Game-based learning presents a uniquely powerful solution to this educational challenge. By simulating realistic scenarios and providing immediate feedback, games can transform abstract entrepreneurial concepts into concrete, engaging experiences. Despite the demonstrated success of game-based learning in fields like sustainability and Privacy, its potential to systematically cultivate entrepreneurship and innovation skills remains largely untapped. Developing and validating game designs for this purpose is therefore crucial for equipping future graduates with the essential competencies to drive economic growth, address societal needs, and thrive in the 21st-century workforce. This research addresses the question: How can a collaborative game be designed to develop entrepreneurship and innovation skills among university students across different disciplines?
Co-supervisor: Dr. Fufen Jin, Engage SFU
The vehicle industry, as well as software and hardware providers are rapidly developing sensor systems and artificial intelligence (AI) methods for sensing the road environment. Connected and Automated Vehicles (CAVs) are argued to have a large potential for accelerating traffic safety and efficiency. Digital twins allow not only to visualize how things work, but also simulate various future scenarios. This is particularly interesting for autonomous vehicles which can be trained in a simulated environment. Furthermore, changes to the algorithm can be validated in a digital twin before deployed on the vehicle. Building a digital twin of a nordic environment allows for development of AI techniques designed for such an environment.
Possible topics:- Gaussian splats: create local environments based on data acquired with an autonomous platform. Dynamic environments that take into account vehicles, pedestrians and cyclists (e.g. MARS, StreetGaussians)
- Underwater splats for representing shipwrecks and other underwater artefacts.
- Digital twins visualization: extend the currently available DigitalTwin of Gløshaugen area and make it more realistic. Final goal is to import it into Nvidia Omniverse so it is usable to train a network that is designed for our autonomous vehicle.- Nvidia Omniverse, CloudXR: visualize a digital twin in VR/AR and simulate various driving conditions
Compared to other instruments such as the piano, the field of automatic processing and generation of guitar music is relatively underdeveloped. This is mainly due to the lack of large, high-quality datasets. The main challenges this project aims to tackle are thus the lack of data and the exploration of Transformer models utilised for automatic guitar tablature transcription and/or for the generation of guitar solos (e.g., in blues). This entails exploring brand-new datasets such as GAPS and SynthTab, and addressing the overfitting to the GuitarSet dataset that is very prevalent in the field, as it is one of the few datasets with a sizeable amount of richly annotated guitar music recordings.
Automatic guitar tablature transcription entails extracting guitar-specific music annotations from pieces of audio recordings of guitar music, while guitar music generation would entail building networks that are able to generate solos having significant variations from the training data and that are capable of capturing long-term dependencies in musical data.
Generativ KI som chat GPT og microsoft co-pilot blir stadig mer brukt i kunnskapsarbeid (som for eksempel i kurs på NTNU). Det er enda et åpent spørsmål hvordan dette påvirker vår evne til kritisk tenking. Kritisk tenkning kan defineres som vår evne til å kritisk vurdere påstander basert på grunnlaget for påstandene. Dette er ansett som en viktig evne i en verden som i økende grad tar i bruk generativ AI med kjente utfordringer som hallusinering, bias etc.
Dette er en empirisk forskningsoppgave. Vi forventer at du bruker rigorøse metoder for å analysere tidligere forskning, planlegge forskningsstudier og gjennomføre empiriske studier.
Du vil gjøre innledende litteraturstudier om emnet og designe en casestudie med datainnsamlingsmetoder som observasjoner, intervjuer og arkivdata.
Det konkrete caset og oppgavens spesifikke fokus vil bli videreutviklet i samarbeid med deg.
Denne oppgaven krever at du har en god forståelse av, og er interessert i, empirisk kvalitativ forskning. Arbeidsspråket for denne oppgaven er norsk. Oppgaven kan skrives på norsk eller engelsk, men vi anbefaler engelsk. Ta kontakt med Marius før du velger denne oppgaven.
This project aims to explore the transformative impact of generative AI on arts by examining how it disrupts traditional processes of artistic creation, audience engagement, and the global art market.
Generative AI will fundamentally disrupt arts. While it might pose a threat to artistic integrity, the uniqueness of artworks, and traditional art market mechanisms, it also affords opportunities to enhance creative processes, the mutability of artworks, and multilateral interactions with the audience. Generative AI can assist in the artist’s creative process through idea generation, style transfer, and real-time experimentation.
The project consists of a literature review in the area which will help narrow down the topic. This will directly impact the empirical part of the thesis, that will also be based on the students’ interests. The second phase includes designing an empirical study and collect data employing qualitative and/or quantitative methods. For example, the project may focus on developing AI tools (e.g., custom get) that support the creation process. In the final phase, the students will analyse the collected data and write up their thesis.
Related work:
Döring, T., Chae, R., Efendić, E., Finken, D., & Nikulina, O. (2026). When Art Meets Algorithm: Exploring How People Perceive Meaning in Human–AI Collaborative Art. Academy of Management Discoveries, (ja).
Rieder, A., Pappas, I. O., & Griffith, T. L. (2025). Trajectories of artificial intelligence: visions from the art world. Journal of Information Technology Case and Application Research, 27(4), 199-206.
To be creative, we need to produce something which is new, meaningful and has some sort of value. Generative AI models such as Transformer-based Large Language Models are able to support humans in creative processes, but to also itself be creative or to assess if an idea or a product is creative. A computational creativity project can investigate any creative field matching the interests and backgrounds of the student or students (language, design, music, art, mathematics, computer programming, etc.), and concentrate on one or several aspects of computational creativity, such as the production, understanding or evaluation of creativity, or on computer systems that support human creativity.
In particular, the project can investigate the transitions between different creative artforms, e.g., generating music or images based on textual input (as in Stable Diffusion models), generating music based on images or text, generating text based on music or images, or generating videos.
Suitable for: One or two students
This thesis focuses on the further development of the ProTuS online learning platform. Students will have flexibility in choosing the exact direction of the work together with the supervisor, for example, within generative AI, self-regulated learning, personalization, or learning analytics. However, the preferred direction is the design, development, and evaluation of a generative-AI-based chatbot that can support learners through conversational assistance, feedback, and guidance.
ProTuS - https://protus.app/, (username: test@usn.no, password: TestTest) is an evolving online learning platform designed to support personalized and engaging learning experiences across different domains. The platform already includes course content, learner interaction tracking, and learning analytics features that can support both students and teachers. Further development of the system opens up several interesting opportunities for research and innovation in educational technology.
One important opportunity is the use of generative AI to provide intelligent feedback, recommendations, conversational assistance, and other forms of adaptive support. In particular, a chatbot integrated into the platform could support learners by answering questions, explaining concepts, guiding navigation, helping with reflection, and providing context-sensitive assistance during learning activities.
At the same time, the thesis remains flexible and can also be oriented toward related topics such as self-regulated learning, personalization, learning analytics, visualization, or other relevant improvements to the platform. The exact scope will be agreed upon together with the supervisor based on the student’s interests and motivation.
The thesis will investigate how an online learning platform can be further developed to better support learners and teachers. The work should be aligned with the ongoing design and enhancement of the ProTuS platform and may address pedagogical, technical, or AI-related challenges.
The exact direction, scope, and implementation details will be agreed upon with the supervisor, based on the student’s interests, motivation, and the platform's needs. The thesis may include the design, development, and evaluation of one or more selected features.
The project consists of the following phases:
Program statement
Safety‑ and security‑critical systems must comply with a range of regulations and standards. However, these regulations are typically expressed as high‑level principles that need to be tailored into concrete checklists for specific systems and application contexts. Traditional approaches rely heavily on human auditors to interpret regulations and develop compliance checklists based on project and system characteristics. This process is costly and requires highly specialized expertise. A further challenge is that a single system may need to comply with multiple standards that can overlap or even conflict. Identifying and analyzing the similarities and differences among these regulations and standards is therefore particularly difficult.
Task descriptions
In this project, we aim to investigate how Generative AI, combined with emerging Retrieval-Augmented Generation (RAG) and Agentic AI technologies, can be used to automate the creation of auditing checklists tailored to a specific system or application, and aligned with one or multiple relevant standards.
The project tasks include: 1) Conducting a literature review of the state of the art; 2) Proposing a Generative AI–based approach and evaluating its applicability, cost efficiency, and usability.
The project will be co‑supervised by Associate Professor Leonardo Montecchi.
Basic knowledge of AI
Git brukes i de fleste programmerings- og prosjektfag på NTNU og andre universitet. I dette prosjektet skal vi undersøke hvordan vi kan monitorerer og automatisere tilbakemeldinger og vurderinger ved hjelp av generativ KI. Målet er en pedagogisk bruk av generativ KI som er godt integrert i arbeidsflyten eks ved bruk av GitHub actions og at faglærere har enkel tilgang til data om prosjektene i egne dashboards.
At the end of their operational lives, ships will eventually become waste that needs to be dismantled properly. Ship dismantling involves various activities, and one of them is inspecting ships to be dismantled. Such inspection is required to ensure the area to be cut does not contain materials and gases that are harmful to workers who will dismantle the ship.
There is an increasing demand for using drones to inspect ships, as drones can reach higher structures and enclosed spaces that are difficult to reach by human inspectors. Ship inspectors usually do not have experience with drone operations, hence there is a need for having graphical user interfaces (GUIs) for remote ship inspections, which are also user friendly for people without any experience using drones.
This project will be carried out as part of the SHEREC project, which aims to improve safety in the ship-breaking process through digitalization and deployment of robots. The students will receive support from the partners affiliated with the project. Multiple students can work on this project provided that they develop different GUIs.
At the end of their operational lives, ships will eventually become waste that needs to be dismantled properly. Ship dismantling involves various activities, and one of them is cutting ship hulls. Currently, hulls are cut manually by workers who use scaffolding or lifted by cranes. The current practice is less safe, as workers are exposed to any accidents that may happen in the cutting area.
There is an increasing demand for using magnetic crawler robots for cutting ship hulls to prevent workers from working at heights. Since cutting workers usually have no experience with robotic systems, there is a need for having graphical user interfaces (GUIs) for operating the magnetic crawler robot, which are also user friendly for people who have no experience with robotic systems.
At the end of their operational lives, ships will eventually become waste that needs to be dismantled properly. Ship dismantling involves various activities, and one of them is cutting ships internally. Currently, internal parts of a ship are cut manually by workers. The current practice is less safe, as workers are exposed to any accidents that may happen in the cutting area.
There is an increasing demand for robotic systems, such as mobile robotic arms, for cutting ships internally so that workers are not exposed to any hazards that may exist in the cutting area. Since cutting workers usually have no experience with robotic systems, there is a need for having graphical user interfaces (GUIs) for operating the mobile robotic arm, which are also user friendly for people who have no experience with robotic systems.
Shipping contributes about 3% of global greenhouse gas (GHG) emissions. Wind-assisted propulsion (WAP) is seen as one of the promising approaches to reduce GHG emissions in the shipping industry, since it allows ships to harness power from wind and reduces fuel consumption. Despite potential benefits of WAP, harnessing power from wind is not a simple task, since the effectiveness of WAP is influenced by multiple operational and environmental factors.
To maximize fuel reduction in WAP ships, there is a strong need to have a system that supports seafarers to harness as much wind energy as possible. In this project, the student will explore graphical user interfaces (GUIs) that will help seafarers maximize power generation from WAP during operations and visualize post-operation information that can be used for later reflections and learnings. The proposed GUIs will also be evaluated with end users to determine their effectiveness.
With the introduction of big-LITTLE or Performance-Efficiency core designs, now present in most modern SoC designs, the industry has started exploring the heterogeneous design space. Yet the predominant approach remains conservative: existing microarchitectures are scaled up or down by adjusting a handful of pipeline structures (e.g., the ROB depth, cache capacity, or issue width), while the underlying microarchitectural design is left unchanged across all cores. The result is that with the currently existing chips, we can only explore a small part of the design space that heterogeneous chips offer, as all the cores still share the same fundamental organization and differ only in resource allocation rather than (micro)-architectural approach.
In this project, we are going to uncover the largely undiscovered region of the heterogeneous design space, where Big-Little cores no longer share the same microarchitecture and are instead free to target vastly different PPA objectives. Rather than continuing to scale the same microarchitectural template, we are exploring co-placing distinct cores, integrating them into a coherent system, and evaluating their combined behavior under realistic workloads.
You will design and implement your own SoC containing at least two architecturally distinct cores, bring up Linux on the resulting system, and benchmark to identify favorable trade-offs across the design space. These three stages are not independent: The system integration work provides the platform, the Linux bring-up validates functional correctness under a realistic software stack, and the benchmark results supply the empirical basis to reason about PPA trade-offs. The goal is to understand how microarchitectural heterogeneity translates into measurable system-level outcomes to address the design challenge that future SoC development poses.
On the technical side, all RTL simulations will be executed on our FPGA infrastructure, using FireSim, a cycle-accurate simulation platform that can achieve near-native execution speeds (if you get Linux to run, you get to interact with the terminal running on your own SoC!). FireSim’s learning curve is steep, and getting hardware up and running is itself a valuable outcome of the project. We expect this project to produce both concrete empirical results and a platform for subsequent exploration of the broader heterogeneous design space.
Further reading
This thesis topic examines how AI systems and broader digital transformation initiatives can be designed, developed, and deployed in ways that prioritize human values and social well-being while ensuring business value. Students can investigate this from various angles (e.g., organizational, technical, or user-focused) and in multiple settings (e.g., healthcare, government, education, or business). Different research methods (e.g., quantitative surveys, qualitative interviews, case studies, or design science) may be employed to explore stakeholder engagement, policy implications, or innovative technical designs.
Multiple students/teams of students can take this topic depending on the interests and skills.
Send me an email explaining why this is interesting/relevant for you.
Related works:
Pappas, I. O., Mikalef, P., Dwivedi, Y. K., Jaccheri, L., & Krogstie, J. (2023). Responsible digital transformation for a sustainable society. Information Systems Frontiers, 25(3), 945-953.
Schmager, S., Pappas, I. O., & Vassilakopoulou, P. (2025). Understanding Human-Centred AI: a review of its defining elements and a research agenda. Behaviour & Information Technology, 1-40.
The field of co-design (also known as participatory design) develops methods and tools to facilitate the inclusion of people of diverse ages, backgrounds, and disciplines in the development of IT products such as smartphone apps, games, and service platforms for citizens and businesses. Participation encompasses all the different stages of the design process: from the analysis of requirements to ideation, prototyping, and technology adoption. Co-design activities were usually performed in the context of in-person workshops facilitated by researchers through the use of physical artifacts (e.g. brainstorming cards). Yet, since the COVID-19 pandemic and the start of work-from-home (WFH) policies, we are now used to hybrid modalities of interaction that heavily leverage digital tools (Zoom, Miro, Teams..). As a consequence, co-design workshops also moved to the digital domain. In this task, we are interested in investigating how to adapt traditional co-design spaces, methods, and toolkits to the hybrid medium and how to rethink interaction among participants. This task will start with performing a literature review on existing work, drafting a simple framework to understand and compare different co-design strategies; and continue developing prototypes of hybrid toolkits. Examples of hybrid toolkits will be provided as case studies.
ICT for Health & Well-being in Built Environments
This project will explore how ICT could contribute to sustainable built environments that support better health and well-being of their occupants. The work will be conducted within the SWELL project: https://www.ntnu.edu/sustainability/swell.
The tasks will include:- A literature review of how ICT could contribute to health and well-being in sustainable built environments.- A literature review of relevant interactive and ubiquitous digital technologies.- Design and prototype of a solutions to engage users of buildings, or other physical spaces.- Evaluation of the prototype.
Properly identifying hate speech is a pressing issue for social media sites as well as for smaller companies, clubs, and organisations that allow for user-generated content. Many such sites currently use slow, manual moderation, which mean that abusive posts will be left online for too long without appropriate action being taken or that content will be published with delay (which might be unacceptable to the users, e.g., in online chat rooms).
The project would look into previous efforts to identify hate speech and cyber bullying, as well as available flame-annotated datasets from chat rooms, online games, Wikipedia, X/Twitter, etc., and investigate ways to identify such language, using various machine learning methods such as Transformer-based techniques and Large Language Models.
The aim of the project is to implement and evaluate a pointer-disambiguation analysis using strict inequalities in the JLM compiler.
Conventional imperative language compilers represent programs internally as static single assignment (SSA) form within a control flow graph (CFG). Although this intermediate representation (IR) is the dominant representation for imperative programs, it bears several drawbacks, such as the SSA maintenance cost, loop (re-)discovery, and the regular loss of important invariants throughout compilation [1]. In contrast, the Regionalized Value State Dependence Graph (RVSDG) is acompiler IR actively developed at NTNU that represents control- and data-flow in one unified representation, avoiding many of the CFGs drawbacks. It is a data-flow centric IR where nodes represent computations, edges represent computational dependencies, and regions capture the hierarchical structure of programs. It represents programs in demand-dependent form, implicitly supports structured control flow, and models entire programs within a single IR.
An important aspect of a compiler is to disambiguate pointers to discover instruction-level parallelism (ILP) such that this ILP can be exploited within optimizations. Several different techniques exist, such as points-to analyses, type-based or flow-based alias analyses, or explicit source code annotations (restrict), to disambiguate pointers and discover ILP. One such technique, namely PointerDisambiguation via Strict Inequalities [2], disambiguates pointers based on the simple observation that two pointers cannot point to the same memory location if one pointer is strictly less than the other pointer. The analysis uses abstract interpretation to build strict less-than relations between pointers.
Currently, the RVSDG is implemented in the JLM compiler [3]. The aim of this project is to add a pointer disambiguation analysis using strict inequalities to JLM and evaluate the implementation in comparison to other, already present, pointer disambiguation analyses. As this project uses cutting edge compiler research tools, a good understanding of compilers and C++ is required. JLM utilizes the LLVM infrastructure, which is commonly used in both commercial and research compilers. This makes this project highly relevant if you are interested in working with compilers in the future. More specifically, the goal of the project is the following:
[1] N. Reissmann, J. C. Meyer, H. Bahmann, and M. Sj¨alander, “RVSDG: An intermediate representation for optimizing compilers,” ACM Transactions on Embedded Computing Systems, vol. 19, pp. 49:1–49:28, Dec. 2020. [Online]. Available: https://doi.org/10.1145/3391902
[2] M. Maalej, V. Paisante, P. Ramos, L. Gonnord, and F. M. Q. a. Pereira, “Pointer disambiguation via strict inequalities,” ser. Proceedings of the 2017 International Symposium on Code Generation and Optimization. IEEE Press, 2017, p. 134–147.
[3] “JLM: A research compiler based on the RVSDG IR,” Mar. 2025,https://github.com/phate/jlm. [Online]. Available: https://github.com/phate/jlm
The aim of the project is to implement and evaluate a global value numbering transformation in the JLM compiler.
Conventional imperative language compilers represent programs internally as static single assignment (SSA) form within a control flow graph (CFG). Although this intermediate representation (IR) is the dominant representation for imperative programs, it bears several drawbacks, such as the SSA maintenance cost, loop (re-)discovery, and the regular loss of important invariants throughout compilation [3]. In contrast, the Regionalized Value State Dependence Graph (RVSDG) is a compiler IR actively developed at NTNU that represents control- and data-flow in one unified representation, avoiding many of the CFGs drawbacks. It is a data-flow centric IR where nodes represent computations, edges represent computational dependencies, and regions capture the hierarchical structure of programs. It represents programs in demand-dependent form, implicitly supports structured control flow, and models entire programs within a single IR. Partial redundancy elimination (PRE) is a compiler transformation that determines when subexpressions are redundant on some, but not necessarily all paths through the program, and eliminates them. It performs a form of common subexpression elimination as well as loop invariant code motion, and for recent formulations based on IRs in SSA form also unifies PRE with global value numbering.
Currently, the RVSDG is implemented in the JLM compiler [1]. The aim of this project is to add a partial redundancy elimination transformation to JLM and evaluate the implementation against the already existing common node elimination transformation. As this project uses cutting edge compiler research tools, a good understanding of compilers and C++ is required. JLM utilizes the LLVM infrastructure, which is commonly used in both commercial and research compilers. This makes this project highly relevant if you are interested in working with compilers in the future.
[1] JLM: A research compiler based on the RVSDG IR, March 2025. https://github.com/phate/jlm.
[2] Karthik Gargi. A sparse algorithm for predicated global value numbering. In Proceedings of the ACM SIGPLAN 2002 conference on Programming language design and implementation, PLDI ’02, pages 45–56, 2002.
[3] Nico Reissmann, Jan Christian Meyer, Helge Bahmann, and Magnus Själander. RVSDG: An intermediate representation for optimizing compilers. ACM Transactions on Embedded Computing Systems, 19:49:1–49:28, December 2020.
[4] Reshma Roy, Sreekala S, and Vineeth Paleri. Partial Redundancy Elimination in Two Iterative Data Flow Analyses. In 38th European Conference on Object-Oriented Programming (ECOOP 2024), volume 313 of Leibniz International Proceedings in Informatics (LIPIcs), pages 35:1–35:19, 2024. ISSN: 1868-8969.
[5] Thomas VanDrunen and Antony L. Hosking. Value-Based Partial Redundancy Elimination. In Compiler Construction, pages 167–184, 2004.
Emerging technologies such as virtual/augmented/extended reality (VR/AR/XR) and artificial intelligence are already transforming how we live, work, and learn. XR has demonstrated strong potential in education by enabling realistic, engaging, and experiential learning environments. At the same time, recent advances in AI offer new possibilities for interaction, feedback, and support in learning processes.
The goal of this master project is to explore how XR combined with ethical, inclusive, and GDPR-compliant AI solutions, hosted locally at NTNU, can contribute to the development of innovative and responsible educational tools.
The project will focus on two main use cases:
VR-based job interview training for NAV. Development and evaluation of immersive VR scenarios for job interview training and career guidance, supported by AI-driven virtual humans. The AI component will be designed with a strong focus on privacy, transparency, and local NTNU-based deployment. The aim is to provide realistic, safe, and scalable training environments for job seekers, while addressing ethical and regulatory requirements. AI-supported study partner for teamwork and collaboration. The project will be performed in collaboration with NAV and potentially Trøndelag fylkeskommune.
Exploration of an AI-based “study partner” that can support students in teamwork contexts, for example within courses such as Experts in Teamwork (EiT).
The system will be designed to facilitate reflection, dialogue, and collaboration, rather than simply providing answers, with particular attention to inclusivity, responsible use of AI, and alignment with pedagogical goals.
Additional cases can be developed depending on student's interests
The project will investigate how such solutions can be designed, implemented, and evaluated in collaboration with relevant stakeholders, including NTNU teachers and external partners. Particular emphasis will be placed on ethical considerations, user experience, and practical applicability in educational and guidance contexts, as well as on the advantages of local, GDPR-compliant AI infrastructures.
The students will have access to a very well-equipped IMTEL VR lab (https://www.ntnu.edu/imtel/) containing Apple Vision Pro, Valve Index, HTC Vive/Vives Pros, Vive Cosmos, 2 Magic Leaps, several Hololenses 1 and 2, Mixed Reality headsets, Oculus Quests, Oculus Rifts, VR treadmill Virtuix Omni & Cyberith Virtualizer, BHaptics suit & gloves, VR laptops etc. A significant number of the VR/AR equipment is portable and can be used at home.
Supervisors: Monica Divitini, Ekaterina Prasolova-Førland (ekaterip@ntnu.no) & Mikhail Fominykh
!!!PLEASE CONTACT Prof. Prasolova-Førland for more information about the task!!!
Advances in extended reality (XR) technologies are opening new possibilities for understanding complex scientific phenomena through immersive and interactive representations. In neuroscience, one of the key challenges is bridging the gap between experimental animal models and human brain structures and functions. This gap often limits the transfer of research insights into clinical practice.
Nevrolens is an XR-based application developed at NTNU in collaboration with the Kavli Institute for Systems Neuroscience and St Olav. It aims to support learning and knowledge transfer in neuroscience by visualising brain structures and processes in an intuitive and immersive way. The next stage of development introduces a novel feature: morphing between species, allowing users to explore correspondences between rodent, macaque, and human brains. This has the potential to improve understanding of how findings from animal studies translate to human biology, with implications for both education and clinical reasoning.
The goal of this master project is to explore how such immersive visualisations can support learning, understanding of neuroanatomy, and transfer of knowledge across species and domains, with potential implications for clinical practice.
Possible research directions include:
Development and evaluation of cross-species XR visualisations
Designing and implementing interactive XR representations that allow users to navigate and morph between rodent, macaque, and human brain structures. The project will investigate how such representations can support conceptual understanding, comparison, and transfer of knowledge.
Pedagogical and clinical relevance of immersive neuroscience tools
Exploring how Nevrolens can be used in educational and clinical training contexts, for example with neuroscience students, medical students, or clinicians. Particular attention will be given to how immersive visualisation can support interpretation of research results and improve translational understanding from animal models to human applications.
The project will be carried out in collaboration with domain experts and will combine technical development with user studies. Emphasis will be placed on usability, scientific accuracy, and the potential of XR technologies to support both learning and future clinical practice.
Supervisors Gabriel Kiss (IDI), Ekaterina Prasolova-Førland (IMTEL/NTNU), Thanh Doan (MH/St Olavs)
Contact: kiss@ntnu.no
Objective: To design and integrate AI-powered chatbots into simulations to provide real-time scaffolding and analyze their effectiveness in addressing learner challenges and improving outcomes.
Description: This research focuses on embedding AI-driven chatbots into Articulate Storyline 360 simulations to provide real-time scaffolding and support for learners. The study will evaluate the effectiveness of these chatbots in addressing common learner challenges, such as navigating complex tasks or understanding difficult concepts. By analyzing learner interactions and outcomes, the research will offer insights into the potential of conversational AI to personalize and enhance online learning experiences.
Create an aquarium (freshwater) that can be monitored via a highly usable web app. Allow users to monitor and interact with the aquarium remotely via a series of sensors and actuators.
The project involves a study of relevant existing research and literature, designing, implementing, and evaluating prototypes (IoT + software), and planning and conducting a series of user tests.
Create a high-tech garden bed that can be monitored via a highly usable web app. Allow users to monitor and interact with the garden bed remotely via a series of sensors and actuators.
Large language models form the basis of almost all currently topical AI research, making it vital to identify whether the models are bias based towards a certain demographic, based on gender, ethical or social background, sexual identity, religion, age, and so on. This has triggered intense research on fair representation in language models, aiming both at building and using unbiased training and evaluation datasets, and at changing the actual learning algorithms themselves. This project could investigate methods to define, identify and quantify a certain bias, as well as develop dibiasing methods, and possibly address under which circumstances a bias in an LLM even could be desirable.
I hvilken grad kan det være mulig og hensiktsmessig å lage en KI-bot for tørrtrening på kommunikasjon med brukere i utforming av brukerhistorier?
I smidig programvareutvikling er brukerhistorier en vanlig måte å representere krav på, og en av mange arbeidsoppgaver man kan komme til å utføre som programvareutvikler / IT-konsulent, er å skrive og kvalitetssikre brukerhistorier i samarbeid med brukere og produkteiere. Den aller beste måten å trene på, er nok i prosjekter med virkelige brukere, men dette er vanskelig å få til å skalere og dermed noe de fleste bare får sjansen til noen få ganger i løpet av en utdanning (f.eks. KPRO / IT2901 ved IDI, og lignende prosjektbaserte kurs andre steder). Noe mer mengdetrening kan man få hvis man har læringsassistenter eller medstudenter som rollespiller som kunder, men dette er også tidkrevende. En mulig alternativ idé er å bruke generativ KI som «bruker» ved å få den til å rollespille i en kommunikasjon (enten med skriftlige prompts, eller også som muntlig konversasjon). Forskningsprosjektet går ut på å undersøke mulige måter dette kan gjøres på, og hvorvidt en slik tørrtrening med KI som «bruker» kan gi nyttig læring av relevante kommunikasjonsferdigheter eller ikke.
Det er i dag mange kunstig intelligenssystemer i bedrifter, alt fra enkle algoritmer, til kompleks bruk av språkmodeller. .
Gjennom denne oppgaven vil man søke å forstå hvordan KI endrer kunnskapsarbeid i praksis. Ett eksempel kan være hvordan KI endrer programvareutvikling.
Du vil gjøre innledende litteraturstudier om emnet og designe en studie som kan inkludere observasjoner og intervjuer.
Oppgaven kan passe for både en og to studenter. Den krever et høyt nivå av selv-styring, og vil være forskningstung. Det spesifikke problemet, og organisasjonen vil bli utarbeidet i samarbeid mellom kandidaten(e), veileder og partnerbedrifter som jobber med denne problematikken. Det er en fordel om du har mulige case du ønsker å jobbe med.
Ta kontakt med Marius før du velger denne oppgaven
I hvilken grad kan en KI-bot spille en troverdig rolle som “elev” med tanke på tørrtrening innen undervisning?
Personer som skal undervise i programmering for første gang, enten det er i rolle som læringsassistent, lærerstudent eller lærer, kan føle seg usikre på rollen og hvordan man bør forklare ting for elevene. Det beste vil nok være mengdetrening på ekte elever, men det har man ikke alltid tilgang til og er vanskelig å få til å skalere med tanke på kjapp øving og tilbakemelding. Spørsmålet vi stiller her, er om det kan være mulig å sette opp et KI-system slik at det kan funke som «tørrtrening» før man skal i gang med ekte elever. Dvs., kan man få KI til å spille en rolle som noen som en nybegynner med begrenset programmeringskunnskap, som gjør typiske nybegynnerfeil med koding og har vanlige misforståelser om hvordan programmeringsmekanismer virker, og hvor en urutinert underviser da kan trene på å gi gradvise hint og forklaringer til KI’en slik at den etter hvert jobber seg fram til en løsning?
I hvilken grad kan en KI-bot være effektiv og hensiktsmessig som “læringsassistent” for nybegynnere i programmering, og hvordan bør den lages?
En utfordring i introkurs i programmering er at det er stort antall studenter og universiteter har kanskje ikke råd til så mange læringsassistenter som det ideelt sett kunne ha vært behov for. Det er imidlertid mye hjelp man kan få fra generativ KI – om enn ikke alltid på den måten som er gunstigst for læring – og noen universiteter legger bevisst opp til at KI kan brukes som LA. Et mål kan da være å få KI til å opptre mer pedagogisk enn å bare gi studenten løsningen med en gang, for eksempel ved å gi hint og hjelp til selvhjelp slik at studenten kan ha større sjanse til å klare å løse et lignende problem på egen hånd neste gang det oppstår. Oppgaven går ut på å utforske bruk av KI til dette formålet, f.eks. lage en prototype på en LA-bot og vurdere kvaliteten på hjelpen den gir – f.eks. med utgangspunkt i et introemne som ITGK.
Det er et enormt etterslep på vedlikehold innen vann og avløp i norske kommuner, et problem som kan bli ytterligere aksentuert med nye EU-regulativer. Oppgaven går på hvordan man kan bruke maskinlæring for å bedre identifisere og prioritere behov for oppgradering i ledningsnettet, og hvordan dette kan inngå i en systemarkitektur for å støtte fagarbeidere i norske kommuner.
Oppgaven struktureres i henhold til Design Science Research (DSR) Problemstillinger utarbeides i samarbeid med Norkart og utvalgte norske kommuner. Ulike versjoner av det som lages skal evalueres i samarbeid med norske kommuner.
Det er å foretrekke at oppgaven besvares på engelsk, men norsk er mulig om det er mest hensiktsmessig. En god besvarelse har resultater som kan danne basis for en vitenskapelig artikkel.
Oppgaven gjøres i sammenheng med Norkart og er reservert for Tobias Kallevik.
Agile methods prescribe practices for development, and were first used in small projects with little criticality. However, such methods are today used also in large projects, and this project will investigate how the practices are adapted and combined with traditional practices to function effectively in large scale where many teams collaborate to develop a large software product. A first generation of large-scale agile methods combined advice from methods such as Scrum with advice from project management. A second generation of methods are currently taken up by the global software industry, with methods such as the Scaled Agile Framework, Large-Scale Scrum, the Spotify model and Disciplined Agile Delivery. The focus on artificial intelligence software and the use of tools for code generation and other activities change both the activities and nature of work.
The project will start as a literature review, but which can be extended with an empirical study for a master thesis project. For an introduction to large-scale agile development, see introduction to special isue in the IEEE Software magazine: arxiv.org/abs/1901.00324 For an example study of a large-scale agile development project see: https://link.springer.com/article/10.1007/s10664-017-9524-2
This assignment will particularly focus on a topic which has received much attention in the literature on large-scale agile development: Coordination of teams. We need to developa better understanding of coordination in large-scale digital product developpment as:• There is currently little advice on inter-team coordination available to practitioners in new frameworks for large-scale agile development such as SAFe and LeSS.• There are few studies of coordination needs over time, and practitioner advice assumes the needs are constant.• Coordination is crucial for the success of complex large Mode 2 IT projects, and more effective ways of coordination will give large societal benefits.• Coordination practices are likely to change due to support of artificial intelligence tools.
A master thesis can be partially based on material from the Agile 2.0 competence-building project where NTNU was a partner, and with other partners DNV GL, Equinor, Kantega, Kongsberg Defence & Aerospace, Sopra Steria and Sticos. The project was led by SINTEF Digital and supported by the Research Council of Norway.
The project work will be done in collaboration with the research project "Rethinking management of large IT projects" (Research Council of Norway) which lasts from 2026 until 2030 and where the primary objective is to "the empirical software engineering discipline by making both theoretical and methodological contributions by rethinking management of large IT projects over time to enable agility at scale". See more information here: https://nva.sikt.no/projects/2763054
If you choose this topic I recommend taking theory modules TDT40 (https://i.ntnu.no/wiki/-/wiki/Norsk/TDT40+-+Advanced+topics+in+software+process+improvement) and TDT39 (https://www.ntnu.no/wiki/spaces/tdt39/overview).
Large-scale digital product development: Learning and unlearning
This assignment will particularly focus on a topic which has received little attention in the literature on large-scale agile development: Learning and unlearning.
Knowledge-sharing is crucial for software project success Software projects that integrate business application domain knowledge with technical knowledge have lower defect density andhigher development efficiency.
We need to develop a better understanding of large-scale learning and unlearning level in digital product development as:• Learning is a key factor to a successful project outcome.• The many practices for learning in large-scale agile methods have not been studied over time.• Practices for unlearning has not been studied in the context of large-scale agile projects.• The practices are likely to change due to support of artificial intelligence tools.
This assignment will particularly focus on a topic close to project management: Portfolio management. Portfolio management focuses on determining how much resources should be spent on which products, and it “provides a global view on resources and their distribution across individual projects according to strategic choices”. We also include strategies for project governance as part of portfolio management, determining type and degree of control in a project or organization. Prior studies suggest that there is often a weak link between agile projects and the organization.
Rapid advances in autonomous and intelligent systems are transforming a wide range of industries, including transportation, maritime operations, aviation, healthcare, logistics, and manufacturing. Technologies such as self-driving vehicles, autonomous ships, delivery drones, and AI-assisted medical systems promise significant improvements in efficiency, safety, and operational performance. However, the successful adoption of these technologies depends not only on their technical capabilities but also on public perception and societal trust. Dimensions such as perceived safety, reliability, transparency, ethical considerations, and overall usefulness strongly influence how individuals evaluate and accept autonomous technologies.
At the same time, public discourse around technological developments increasingly takes place on online platforms, particularly through video content. Platforms such as YouTube host thousands of expert discussions, industry presentations, technology demonstrations, and public commentary related to emerging autonomous systems. These videos represent a rich but largely unstructured source of information about how autonomous technologies are discussed, framed, and perceived by experts, journalists, and the broader public.
With the exponential growth of online video content, extracting and analysing spoken information from videos has become an increasingly valuable task. The goal of this thesis should be to develop an application that automates the retrieval and analysis of transcripts from YouTube videos, particularly from curated playlists focusing on autonomous technologies across multiple industries. By leveraging large language models (LLMs), the system will categorize, summarize, and analyse discussions related to key dimensions such as safety, reliability, trust, efficiency, and societal impact. The aggregated analysis will allow the identification of patterns and trends in how advances in autonomous systems are communicated and perceived over time.
By combining automated transcript retrieval, natural language processing, and large language models, this thesis aims to provide a scalable framework for analysing large volumes of spoken online content. The resulting system will enable researchers and practitioners to better understand how autonomous technologies are represented in public discourse and how perceptions evolve as these technologies mature.
As a first step, students will conduct a literature review to explore existing methodologies for transcript extraction, natural language processing, and topic modelling using large language models. The next phase involves designing and implementing a system that fetches transcripts from YouTube videos via API, organizes them, and applies LLM-based techniques for data analysis.
The system should enable:
Upon implementation, a user study will be conducted to evaluate the system's effectiveness in organizing and presenting information. The thesis will conclude with an analysis of the collected data and a discussion on future improvements.
Candidates should have a background in software development and an interest in AI-driven data analysis. Essential skills include:
A software product line (SPL) is a set of software-intensive systems sharing a common, managed set of features that satisfy the specific needs of a particular market [1]. Using some Variability Realization Mechanism (VRM), a SPL encodes a potentially large variety of software products as variants of some common code base. A typical concrete example of a VRM is conditional compilation in C/C++, realized through pre-processor directives (#def, #ifdef, etc.); such technique is for example used in the Linux Kernel, which can be considered a prominent example of SPL.
While SPL streamline the maintenance of similar products from a unique code base, they also introduce several challenges, for example an increase in complexity in handling corner cases, since any change to a single feature will be reflected on all the products of the product line. This project aims to understand how LLMs can help software developers in effectively managing SPLs and their evolution, building on recent work from the literature [2].
Mobile phones are actively used to access the weather and weather forecast information. For example, the Norwegian Meteorological Institute maintains access to its weather services and data in the app Yr. Much less attention has been paid for presentation of climate information, specifically when local climate information is to be considered. The climate information, however, plays an important role in human decision-making, guiding agricultural and construction activity and long-term planning.
The objective of this project is to design such a climate information app for mobile phones.
Deliverables. The app shall overlap the static geographical and geomorphological information (from, e.g., Open Street Map) with ecological and climate information from high-resolution online datasets. The datasets are provided by, e.g., the ESA CCI COPERNICUS projects. For ecological information, one may take ESA CCI for land cover or CORINA dataset. For climate information, E-OBS or gridded high-resolution land surface temperature information will fit. The final set of climate and geographical information data sources will depend on the design of the app and must be specified in a dialog with the supervisor.
This work will be conducted within the International project URSA MAOR: https://www.sintef.no/en/projects/2021/urban-sustainability-in-action-multi-disciplinary-approach-through-jointly-organized-research-schools/
Læringsteknologi er programvare og andre teknologiske produkter som understøtter læring og undervisning. Her er det mulighet for selvvalgte oppgaver enten fra studenter eller studenter i samarbeid med fagstab, og prosjekter som kan relateres til enten Excited senter for fremragende utdanning.
The Challenge: Fighting Invisible Threats with Computer ScienceEvery year, over 4.3 million patients in the EU acquire Healthcare-Associated Infections (HAIs). A major cause is invisible airborne pathogens disrupting the sterile airflow in Operating Rooms (ORs) during surgery. While Computational Fluid Dynamics (CFD) can simulate these risks with high fidelity, it takes supercomputers days to process millions of data points. Surgeons cannot see this data, meaning they often unknowingly break the protective air curtain.
Our Mission: The Delivery Truck for Real-Time Physics at NTNU’s Department of Computer Science (IDI) and St. Olav’s Hospital, we are building an Extended Reality (XR) framework to solve this. As part of the EU Horizon Europe HumanIC project, we are developing the software pipeline that translates heavy engineering simulations into a real-time, interactive "sixth sense" for surgeons, aiming to reduce infection risks by 30% and increasing energy efficiency by 10%.
The Tech Stack You Will Work With: If you join our research group, you will tackle some of the hardest problems in modern computer graphics and AI:•Next-Gen Digital Twins: Traditional photogrammetry fails on the shiny, reflective stainless steel and glass of medical equipment. We are pushing the boundaries of 3D Gaussian Splatting (3DGS) to capture hyper-realistic, geometrically accurate OR environments instantly using consumer cameras and smartphones.•Lagrangian-Based Middleware: CFD software (Ansys) and game engines (Unreal) do not speak the same language. We are building custom Python and Blender middleware that extracts raw particle tracking data and "bakes" the physics (velocity, particle age, temperature) directly into mesh vertex attributes.•High-Performance XR Rendering: We are developing hybrid rendering pipelines that combine lightweight 3D meshes for room architecture with photorealistic 3DGS for medical tools, achieving a motion-sickness-free 72+ FPS on PC-VR (Meta Quest 3 + RTX 4090).•Real-Time Edge AI: To make airflow predictions instant, we are bypassing traditional physics solvers. We train AI Surrogate Models (Neural Operators) to act as a "SmartInbetweener," predicting chaotic, non-linear turbulence in just 5-10 milliseconds directly on the GPU’s Tensor Cores.
Suggested research topics :1.Tackling "Phantom Geometry" from Reflective OR Equipment in 3DGS (Computer Vision)2.Enhancing SfM Point Could initialization in Texture-less Surgical Environments (Image Processing)3.Cognitive Ergonomics and UI/UX Evaluation for Airflow Visualization in AR (HCI)4.Synthetic Dataset Generation for Medical Digital Twins (Data Engineering)5.Hardware-Aware AI Surrogate Modeling: Unlocking Asynchronous Compute and FP8 Inference for Real-Time CFD (ML-Computer Architecture)
Supervisors: Gabriel Kiss kiss@ntnu.no Rahmat Rizal Andhi rahmatra@stud.ntnu.no Frank Lindseth frankl@ntnu.no (COMP/IDI)
Artificial Intelligence (AI) is reshaping the way software is designed, implemented, and maintained [1]. Modern developers are exposed to many new topics and concepts, which are also evolving rapidly. Concepts like prompt optimizaiton, context windows, RAG, model size, fine-tuning, etc., have all a great impact on the quality attribute of the final system, for example on performance and reliability, or even on costs, if using commercial models. AI-based components, and in particular LLMs, are being increasingly integrated into software systems, and the impact of Generaitve AI on software development is expected to be long-term and multi-faceted [2][3].
This project aims to identify a comprehensive set of AI‑related concepts that are relevant to professionals in software engineering (e.g., MLOps, prompt engineering, generative code synthesis, prompt optimization, etc.), and to map them to a Software Engineering (SE) teaching curriculum [4]. The objective is to systematize the different initiatives and provide a reference curriculum covering both SE4AI (Software Engineering for AI) and AI4SE (AI for Software Engineering) aspects.
The project aims to study various aspects of learning to program using biometric sensors such as EEG (brain activity), eye tracking (gaze and attention), and GSR (galvanic skin response) sensors. Potential scenarios could be comparing tasks with and without AI assistance for example.
The project involves a study of relevant existing research and literature, planning and conducting a series of user tests. Furthermore, it is expected that such a test will generate a wealth of data to be analyzed and interpreted to draw out interesting and useful results and conclusions. Depending on the case it might be necessary that the students develop data processing scripts and or novel visualizations.
This is a work in progress.
Multiple projects will be presented. Each is available for one or two students, either as a specialization project or a master’s project.
Description
Together with MIRA, SINTEF Health, the medical faculty at NTNU and St Olav university hospital we are offering medical image computing (MIC) projects, based on Deep Learning (DL) and Computer Vision (CV), and related to:
Some concrete examples:
It's also possible to focus more on model dev. E.g. U-mamba, that can be tested on various organs and modalities.
It's desirable to use MONAI (and MONAI Label) for all developments. We encourage and help to publish papers based on some of the master-thesis work done. Several PhD students work on the topics so you will have extensive help during the master.
You can work individually, or in pairs of two.
Modeling and documenting the software architecture [1] is a fundamental task in software engineering, and established modeling languages (e.g., UML) have been used for this purpose. Different organization of components and patterns [2] have a large impact of non-functional attributes of a system, such as reliability, security, performance, etc. Models, as communication mechanism, have been fundamental in popularizing design patterns for classical software systems [2]. This project aims to investigate languages and patterns for modeling software architectures that include AI components.
One important tool in software architecture specification are modeling languages, and in particular Architecture Description Languages (ADLs). The building blocks of an architectural description are: 1) components, 2) connectors, and 3) architectural configurations. An ADL must provide the means for the explicit specification of those aspects [3]. Traditional examples of ADL include UML, SysML and AADL, and the recently released SysML 2.0 [4].
The objective of this project is to investigate the limitation of those languages for what concerns the modeling of software architectures and design patterns for AI components. For example, concepts like AI models, prompt patterns, AI agents, fine tuning, etc., are not explicitly captured by traditional ADL.
Multiple projects will be presented. Each is available for one or two students, both as a specialization project and a master’s project.
Cranes are traditionally controlled by operators who work inside the crane’s cabin. Although this operation mode is still common nowadays, a significant amount of progress has been made to move operators away from their cranes, so they would not be exposed to hazardous situations that may occur in their workplace. Although the transition to remote operation brings many benefits, the amount of sensorial information is significantly reduced in remote crane operation, which may reduce operators’ capability to work safely and productively. Therefore, it is important to investigate how multimodal user interfaces could be used to bring back some of the missing sensorial information.
Key activities in the project include:
This project is part of the OpenRemote project, which aims to develop open-source design systems for remotely operated machines. The student will receive support from the partners affiliated with the project. Multiple students can work on this project provided that they develop different multimodal user interfaces.
Game development is a large well-known area in traditional web development. However, it is still to be seen how the emerging web3 technology will take it a step further!
In this project, a multiplayer game will be developed using a mix of traditional web and modern web3 technologies! It will be an excellent opportunity for students to learn about new technologies and possibly apply those later in their thesis or in their further careers.
The game that will be developed will be a multiplayer Blackjack game unless there is a better proposal from students! Technically, it will have following components
The game will have following basic functionalities
Research Aspects
Open data involves the pooling and collecting of data across a community, industry or group of stakeholders. The motivation is the vision (aspiration, hope, belief...) that by making data openly availble, hence accessible to everyone, this will boost productivity through enhanced collaboration or create more well-functioning markets. Examples include: Open Target in pharmaceutical industry, the EU's PSD2 regulative towards open banking in finance, or HUNT research database at NTNU.
Visions of the role of open data to are widespread as illustrated by this recent Stortingsmelding, https://www.regjeringen.no/no/dokumenter/meld.-st.-22-20202021/id2841118/?ch=5
The challenge, however, is that the mere availability of open data is not sufficient for its uptake and use towards collaboration. There are social, practical and institutional conditions that need to be in place for visions of open data to materialize.
The student(s) will analyse this for a particular proposal for open data, Open Data Subsurface Universe (https://osduforum.org/). This is a data platform for sharing, communicating and doing analytics of data. It orgininated and has a foothold in the fossil energy sector, but is moving into renewable energy and CCS installations too as OSDU is a general framework for capturing any physical, geo-located asset (similar to Digital Twins).
The students will do their projects with partner companies. Presently there are two: Equinor and AkerBP.
If you are interested in the areas of Child-Computer Interaction, Educational Technology, Project-based learning, Learning technologies etc. you can contact me at spapav@ntnu.no to discuss on a topic, adequate for a Master’s project.
Medstudentvurderinger hvor studenter gir tilbakemeldinger på hverandres innleveringer brukes mye som læringsaktivitet. En av utfordringene er at tilbakemeldinger kan være av variabel kvalitet, motstridende og med mange som git tilbakemelding blir det mye å se over. I dette prosjektet skal vi se på løsninger for oppsummering av medstudentvurderinger. Oppgaven bygger videre på arbeid som er utført i tidligere oppgaver og tema for fordypningsprosjekt kan være å prøve ut systemet i et faktisk emne. Som masterprosjekt er det eksemplevis mulig å se på forskjellige presentasjoner av positive og negative kommentarer og mekanismer for å gi tilbakemelding på tilbakemeldingene.
Artificial Intelligence is now being used at an increasing rate to augment or automate organizational decision-making. From processes such as performing credit checks on customers of banks, aiding in forecasting of future events, and automating manual and repetitive tasks, AI is introducing a new way of making decisions for organizations. The purpose of this project is to examine through empirical methods the effects and processes of transition to AI-based decision-making structures.
Work on an interesting project related to orientation sensing detection (device relative to the user) in order to provide accurate audio instructions to blind people, for example.
The project involves a study of relevant existing research and literature, designing, implementing, and evaluating prototypes, and planning and conducting a series of user tests.
Cranes are traditionally controlled by operators who work inside the crane’s cabin. Although this operation mode is still common nowadays, a significant amount of progress has been made to move operators away from their cranes, so they would not be exposed to hazardous situations that may occur in their workplace.
Typical setups for remote crane operation consist of multiple monitors, where video streams and graphical user interfaces (GUIs) are presented on separate monitors to prevent the GUIs from blocking the video streams. However, this approach provides its own challenge, since operators must split their attention and manually integrate multiple pieces of information located at different places. Therefore, there is a need to design suitable overlaid visual information, where graphical information can be overlaid on video streams without blocking important objects.
This project is part of the OpenRemote project, which aims to develop open-source design systems for remotely operated machines. The student will receive support from the partners affiliated with the project.
What is the most natural way to get information about bus scedules or other well organised and structured data?
In Trondheim we have the bus oracle (BusTUC) which has answered tens of millions English and Norwegian questions about the buses in Trondheim since 1997.
You can try it yourself at http://busstuc2.idi.ntnu.no
On the other hand, there are also several services that answer many of the same questions, just by selecting location, destination and time constraints.
The goal is to investigate the best way for different groups of users, in different situations, to get travel information. The work can follow up the FUIROS project, investigating Speech, GPS, Maps, or Real-time data integration and route updates.
One example problem that could be investigated is that when people ask the Oracle and get an answer they are not happy with, they have to start from scratch again, because the system "forgets everything" between questions. How can we make the Oracle react appropriately to each single user's needs, by remembering everything that they have asked before? How does that fit within GDPR regulations?
Some existing work to build on can be found here.
Many regular maintenance operations occur over the lifetime of a commercial building. This includes for example replacement of air filters which filter the air supplied into a building. Short maintenance cycles stay on the safe side by replacing filters too often before any efficiency loss or down-time occurs. This may lead to time and material consuming replacements before they are actually necessary.
In an initial step, promising regular maintenance operations for automated prediction need to be identified and ranked based on their economic impact.
The goal of this thesis is to develop predictive maintenance methods for one or multiple of the identified operations in order to reliably detect the need for replacement or maintenance before a problem occurs.
This project is in collaboration with Piscada, a Trondheim-based technology company that develops an industrial cloud-based software platform for customers in construction and energy (PropTech), Industrial IoT, aquaculture, and general process management. The company was established in 2009 as a spin-off from SINTEF and focuses on innovation and simplification of industrial IT systems, as well as building a bridge between industrial automation and IT. There are today approximately 2,000 installations of Piscada's software and a diverse list of renown customers. We aim to be a leading industrial service platform with a focus on effective monitoring, new insights and optimization for increased sustainability in selected industries.
Many regular maintenance operations occur over the lifetime of a fish farm. This includes for example cleaning of the feeding mechanism or the tubes through which the feed is distributed to the fish-nets. Short maintenance cycles stay on the safe side by cleaning too often before any down-time or damage occurs. This may lead to time-consuming cleaning before it is actually necessary. Many fish-farm operators develop a good intuition for when a cleaning cycle is necessary, but this is not easily reproducible or transferable across employees.
In an initial step, promising regular maintenance operations for automated prediction need to be identified and ranked based on their economic impact. The goal of this thesis is to develop predictive maintenance methods for one or multiple of the identified operations in order to reliably detect the need for maintenance before a problem occurs.
This project is in colalboration with Piscada, a Trondheim-based technology company that develops an industrial cloud-based software platform for customers in construction and energy (PropTech), Industrial IoT, aquaculture, and general process management. The company was established in 2009 as a spin-off from SINTEF and focuses on innovation and simplification of industrial IT systems, as well as building a bridge between industrial automation and IT. There are today approximately 2,000 installations of Piscada's software and a diverse list of renown customers. We aim to be a leading industrial service platform with a focus on effective monitoring, new insights and optimization for increased sustainability in selected industries.
As per www.regjeringen.no, zoning plans specify the use, conservation and design of specific geographical locations. They consist of detailed land-use plan maps that are coupled with a planning provision and plan description. When looking to start a construction process in a given area, reviewing the corresponding zoning plan is essential. This is where one can find information regarding factors such as where in the area buildings can be placed, as well as certain characteristics (ex: height, roof style) the buildings must abide to. Accessing and understanding the zoning plans, however, can be a complex and time-consuming process for citizens, developers, and even case workers. Therefore, citizens and developers often rely on contacting municipal offices directly for explanations and guidance, which can be inefficient and time-consuming for both parties. It is therefore in the best interest of the municipalities of Norway that a solution for easy retrieval of information from zoning plans is developed.One such solution, “Planslurpen,” is part of DiBKs “Drømmeplan”-project, and the end goal is for it to be a national component available to everyone. It uses machine learning methods to retrieve key information from zoning plans and presents it in a manner that allows one to easily find which regulations apply to a chosen area. It is not ready for deployment yet, though. For example, currently, the plan-id and plan description must be manually specified and uploaded, which would not be ideal in production. High quality data flow and output are key factors in determining the success of Planslurpen.In this project, the students will be working closely with the municipalities of Trondheim and Kristiansand, stakeholders such as DiBK and KS, and the developers of Planslurpen. The project has a high degree of freedom, as the students will assess the needs of all involved parties and contribute to the further development of Planslurpen based on their findings. Potential approaches could include designing a data infrastructure for easy integration of Planslurpen in municipal processes, development of multi-agent AI chatbot functionality, suggestions for improvement of the Planslurpen API, or researching methods to improve Planslurpens retrieval and presentation of zoning plan details.Throughout the project period, the students will have access to expert competence in the field of zoning plan case handling from the municipalities of Trondheim and Kristiansand, for informative and testing purposes. They will also be working with DiBK, KS and the developers of Planslurpen. The students will have access to raw data from the municipal zoning plan registries for the Trondheim and Kristiansand municipalities, which consists of several thousands data points. Data will also possibly include the data used to train Planslurpen, although this is yet to be confirmed. It will likely be confirmed by the end of March.
Project thesis outline and objectives: Develop an understanding of the problem space Discern the needs of involved parties Evaluate the current Planslurpen architecture and data flow Explore potential approaches Literature review covering state-of-the-art methodsExample objectives for master’s thesis: Development implementation of multi-agentic AI architecture for a zoning plan chatbot Proof-of-concept implementation of AI-friendly Planslurpen API optimizations Development of scalable and interoperable architecture for integration into other municipalities Evaluation of proposed ideas through continuous dialogue with stakeholders Development and implementation of methods to improve Planslurpen Increasing user trustworthiness of Planslurpen through explainability
Video:
DNV is currently leading a project under the auspices of ESA (European Space Agency) that focuses on the use of satellite data within shipping in the Arctic and Baltic Sea regions. The project aims to identify the needs for various types of satellite data, which services and products currently offering this, the extent and in which manner the satellite data is being used, and similar aspects. The current work on this project is published as reports on https://earsc-portal.eu/display/EO4BAS. The EO4BAS project is part of a larger project within EO data (Earth Observation, i.e., satellite data) financed by ESA and EC (European Commission). Not only opportunities within the maritime are explored, but also within ex. oil and gas, and raw material extraction.
MLIR is emerging as new trend to solve a fair number of issues in the compilation pipeline by gradually lowering abstractions into equivalent formulations closer to the hardware. Despite very promising case studies, the overall infrastructure is less mature than competing technologies. This project focus on one such areas where the maturity of the framework is lacking with respect to more traditional alternatives, e.g., LLVM.
We present the case for MLIR code instrumentation to be developed into profile-guided optimisations required in an ongoing compiler construction effort by researchers in NTNU.
The Data Recomputation for Multithreaded Applications [1] paper demonstrated that recomputing values from their defining expressions — rather than buffering and reloading them — can reduce memory traffic and improve cache utilisation in parallel workloads. This project is meant to bolster an ongoing project on MLIR-level recomputation and memory fission optimisations, which currently rely on a static cost model to make these decisions. This project will explore how profiling data can sharpen these decisions, potentially leading to better performance.
Both passes currently implement a static cost model: recomputation is chosen over buffering when an estimated numConsumers × computeCost beats computeCost + numConsumers × loadLatency, where load latency is inferred from a three-level cache hierarchy (L1/L2/L3/DRAM). The model, however, has known limitations — trip counts are unknown statically, buffer residency in cache is assumed rather than measured, and per-operation cycle costs are generic defaults loaded from a JSON file. These approximations can flip the optimal decision in either direction when poorly tuned.
The following plan covers the expected work directed towards a master thesis, while the pre-project should focus only on a subset of these aspects, to be agreed with the supervision team.
The student will investigate how profiling data can sharpen these cost decisions at the MLIR level [2]. Concretely:
Instrumentation pass — Insert lightweight counters at loop entry and load/store sites in MLIR, lowerable through the standard pipeline to executable code. This mirrors the block-frequency infrastructure used in LLVM's profile-guided optimisation (PGO), adapted to operate before lowering where loop structure and memref shapes are still explicit.
Profile ingestion — Parse profiled edge/block frequencies and attach them as MLIR attribute metadata (or as affine.for trip-count annotations). Evaluate whether mlir-opt's existing --mlir-print-ir-after-all and diagnostic infrastructure can surface this cheaply.
Cost model integration — Feed real trip counts and dynamic load-miss rates into decideBufferStrategy() and the MemoryFission candidate scoring, replacing the static fallbacks (e.g., the current hardcoded L1-residency assumption in MemoryFission).
Evaluation — Measure decision accuracy and wall-clock improvement on a set of microbenchmarks and real-world kernels (e.g., from PolyBench or ML workloads) that are amenable to recomputation. Compare against the static model and an ideal oracle that has perfect knowledge of dynamic behavior.
The project will produce an instrumented MLIR pipeline, a profile-driven cost model, and a quantitative evaluation.
[1] G. G. Akbulut, M. T. Kandemir, M. Karakoy and W. Choi, "Data Recomputation for Multithreaded Applications," 2023 IEEE/ACM International Conference on Computer Aided Design (ICCAD), San Francisco, CA, USA, 2023, pp. 01-09, doi: 10.1109/ICCAD57390.2023.10323776.[2] C. Lattner et al., "MLIR: Scaling Compiler Infrastructure for Domain Specific Computation," 2021 IEEE/ACM International Symposium on Code Generation and Optimization (CGO), Seoul, Korea (South), 2021, pp. 2-14, doi: 10.1109/CGO51591.2021.9370308.
Digital transformation is influencing all the workplaces. Not always the digital transformation that is envisioned is successful, as witnessed by, for example, the challenges connected to the introduction of the Helseplatformen. One aspect that is often under-estimated is connected to the competences that are needed to workers to participate to the digital transformation in a meaningful way.
This task aims at designing a toolkit/game to help workers to understand the space of possibility of new technologies in their workplace. Focus will be on promoting creativity and system-thinking, at the same time keeping into account the constraints that are given by the existing infrastructures.
Contact the supervisor to share your ideas and know more about this task.
This topic is about creating a self-learning multi-agent scheme for steering simulated cars in an urban environment. The focus is on developing a Reinforcement Learning scheme for this application.
The project builds on existing results along the lines sketched above, and can be focused on different aspects, depending on the interests and the pre-knowledge of the student(s).
One possible focus is to perform the research in the well-known realistic autonomous driving simulator CARLA. Another possibility is to focus the investigation on using existing small floor robot platforms that we have, that is the LIMO robots by Agilex.ai
Research on recommender systems has seen thousands of studies being published over the course of the last two decades. Frequently, author report that their proposed method performs better than the state-of-the-art.
In reality, evaluation design features a multitude of choices which leads us to question whether new methods deliver the claimed added value. Examples for these choices include:
The master project will look into a specific genre of recommendation: news. There, mulitple data sets exist that allow researchers to assess how well different recommendation algorithms perform. The candidate will carry out a set of experiments to determine how reliable are published results.
The notion of responsible AI entails a large range of aspects regarding how AI applications are developed, utilized, and monitored throuhgout their lifecycle. The purpose of this project is to explore what responsbile AI means for organizations, which processes and structures they are establishing in order to attain set indicators of responsible AI, as well as what are the organizational impacts of it. Does adopting responsbiel AI result in any organizational gains? Does it influence how customers/citizens perceive the organization, or is it restricting what they can do with novel technologies?
Please contact me before choosing this task.
Responsible digital transformation is to understand the concept of responsibility as it pertains to the enablement of positive outcomes of ongoing digitalization of our everyday lives while safeguarding the human experience against possible negative consequences of the same.
Not all digital transformation is positive. Take social media for example. Many now claim that social media not only correlates with increased psychological illness, but is the reason for it.
From a sociotechnical perspective, we assume that humans are able to influence technology, or at least, restrict how it impacts us by not using it. With social media for example, we must question such assumptions. Can we realy choose not to be on social media, or would we then be effectively excluded from society? One way to approach the issue of responsibility would be to consider if and how digital tools help humans achieve freedom to act in the world. For example, act to pursue sustainability goals.
In this project you will be given the opportunity to explore the concept of responsible digital transformation. It is good if you have a case in mind that you want to explore.
The project will start as a literature review, but can be extended with an empirical study for a master thesis project.
The task can be taken by one or two students. The task requires a high level of self-management and is research-intensive. Knowledge of research methods is important (e.g. IT3010 or module TDT39). The specific case and the focus for the task will be developed through a dialog between the candidates and the supervisor.
Enhetlig pasientbehandling blant helsearbeidere i helsesektoren undergraves av ulike former for grenser - eografiske, institusjonelle og profesjonelle. Dette er til hinder for effektiv og høykvalitet pasientbehandling. Eksempler inkluderer samarbeide mellom fastlege og sykehus, eller samarbeide mellom sykehus og kommunehelsetjenesten herunder eldreomsorgen.
Trass pådriv og initiativ for å få helsearbeidere til å samarbeide tettere og mer interaktivt, gjenstår mye. Informasjonssystemene i helsesektoren er "silo"-orientert dvs de understøtter primært arbeidet lokalt, ikke samarbeide gjennom behandlingskjeden.
Det har opp gjennom årene vært satt i gang en rekke reformer og tiltak (feks Samhandlingsreformen, En innbygger en journal, Helseplattformen) uten at dette har løst utfordringene.
IKT (digitalisering) blir pekt ut som mulig løsning, gitt kapasitet til støtte distribuerte arbeidsprosesser.
Prosjektet/ oppgaven vil ta for seg et utvalgt innføringsløp for en digital tjeneste i helsesektoren. Oppgaven vil innebære en selvstending, empirisk innhenting av krav gjennom observasjon, intervju og logging av bruk av eksisterende system. Krav/ behov skal så operasjonaliseres i anbefalte, ev også prototypet, funksjonalitet.
The project aims to study various aspects of creating a solution that facilitates sharing office/desk use, converting them into “smart” desks or “context-aware” desks.
Moderne krigføring er høyteknologisk med bruk av droner og KI. Dette krever raske innovasjonssykluser der droner og systemer må tilpasse seg raskt. I tillegg brukes og tilpasses standardkomponenter som gjør at utvikling av slike systemer kan skje utenfor de tradisjonelle, store leverandørene. Spørsmålet er hvor godt Norge, norske leverandører og det norske forsvaret er forberedt på denne typen rask og smidig utvikling av teknologi.
I denne oppgaven vil vi søke å få 1) oversikt over hva forskningen sier om beste praksis for smidig og tilpasningsdyktig utvikling av droneteknologi, og 2) undersøke med norske forsvarsstartups og dronefirma hva deres perspektiv på denne typen utviklingen er.
Det konkrete caset og oppgavens spesifikke fokus vil bli videreutviklet i samarbeid med deg. Det er en fordel om du har kontakter med forsvaret, dronefirma eller liknende og er interessert i å jobbe med dette feltet.
Denne oppgaven krever at du har en god forståelse av, og er interessert i, empirisk kvalitativ forskning. Arbeidsspråket for denne oppgaven er norsk. Oppgaven kan skrives på norsk eller engelsk, men vi anbefaler engelsk.
Ta kontakt med Marius før du velger denne oppgaven.
The goal of this project is to explore the application of models commonly used for predicting species discovery to the task of identifying vulnerabilities in software systems. Drawing parallels between the process of species discovery and software vulnerability detection, the proposal is to develop or adapt models inspired by species accumulation curves to analyse the cumulative number of software vulnerabilities discovered over time. By utilising historical vulnerability data and considering factors such as software complexity and codebase size (if available), these models will seek to predict the rate of new vulnerability discoveries and estimate the total number of vulnerabilities within a software system. Experience with statistical models and methods will be instrumental.
The research outcome would aid software vendors and security researchers estimate the number of vulnerabilities in software based on historical data, as well as adapt new strategies for bug hunting when current methods predict a low number of future discoveries.
https://www.youtube.com/watch?v=TeY1fY0Bi_M
https://scholar.google.co.uk/citations?view_op=view_citation&hl=de&user=bX-GbkUAAAAJ&citation_for_view=bX-GbkUAAAAJ:hFOr9nPyWt4C
The aim of the project is to implement and evaluate a static high-level synthesis back-end in the JLM compiler.
Conventional high-level synthesis (HLS) tools represent programs internally as static single assignment (SSA) form within a control-data flow graph (CFDG). Although this intermediate representation (IR) is the dominant representation for high-level synthesis tools, it bears several drawbacks, such as the SSA maintenance cost, loop (re-)discovery, and the regular loss of important invariants throughout compilation [1, 2]. In contrast, the Regionalized Value State Dependence Graph (RVSDG) with its R-HLS dialect [2] is a compiler IR actively developed at NTNU that represents control- and data-flow in one unified representation, avoiding many of the CFDGs drawbacks. It is a data-flow centric IR where nodes represent computations, edges represent computational dependencies, and regions capture the hierarchical structure of programs. It represents programs in demand-dependent form, implicitly supports structured control flow, and models entire programswithin a single IR.
The current implementation of the R-HLS dialect supports a dynamic handshake protocol to synchronize program operations when synthesizing the corresponding hardware. This handshake protocol is a good fit for programs that are affected by dynamic behavior, such as cache misses or unpredictable branches, but results often in overhead for programs that experience no such behavior. An alternative to a dynamic handshake protocol is a statically created state machine that controls program execution of the synthesized hardware.
Currently, the R-HLS dialect is implemented in the JLM compiler [3]. This project aims to add a static HLS back-end and evaluate it in comparison to the dynamic back-end. As this project uses cutting-edge compiler and HLS research tools, a good understanding of compilers, C++, and a basic understanding of digital hardware is required. JLM utilizes the LLVM infrastructure, which is commonly used in both commercial and research compilers. This makes this project highly relevant if you are interested in working with compilers in the future. More specifically, the goal of the project is the following:
[2] D. Metz, N. Reissmann, and M. Sj¨alander, “R-HLS: An IR for dynamic high-level synthesis and memory disambiguation based on regions and state edges,” in Proceedings of the 43rd IEEE/ACM International Conference on Computer-Aided Design, ser. ICCAD ’24. [Online]. Available: https://doi.org/10.1145/3676536.3676671
Immersive technologies such as virtual/augmented/extended reality (VR/AR/XR) have demonstrated significant potential in transforming educational practices by providing learners with realistic and highly engaging learning experiences. In most cases, due to budget and practical concerns, educators use relatively unexpensive XR equipment such as Oculus Quest. While this might be sufficient for many educational situations, it is important to investigate the potentials of more advanced equipment that provides advanced spatial computing possibilities, simulates senses other that sight and hearing and facilitates walking.
The goal of this project is to explore how advanced XR technology, beyond the regular XR equipment, could support learning, especially at university and professional level. The specific topic of the project will be defined in collaboration with the student depending on the choice of equipment. Here are examples of possible projects:
Supervisors: Monica Divitini, Ekaterina Prasolova-Førland (ekaterip@ntnu.no) & Mikhail Fominykh!!!PLEASE CONTACT Prof. Prasolova-Førland for more information about the task!!!
This thesis focuses on designing and implementing a digital toolkit that supports, guides, collects data and in general aims to offer a fruitful learning experience for young students (K-12). Young students engage in learning when they take part in collaborative problem-solving experiences allowing them to generate inventive solutions to real-world issues. This experience can be supported with the help of digital tools that can also leverage the use of AI.
The initial stage of the candidate(s)’ work will involve conducting a comprehensive review of existing literature on how current digital toolkits support learning processes. Based on this review, the candidate(s) will define their contribution by identifying gaps, best practices, and specific requirements that can drive innovation in the proposed toolkit.
The toolkit—implemented as a platform, website, or app—will be designed using an iterative approach, with parallel efforts focused on data collection, storage, and analysis. Subsequently, the candidate(s) will conduct an on-site testing session with target users to evaluate the effectiveness of the proposed toolkit and gather data.
Finally, the candidate(s) will analyze the results to complete the design process and establish the foundation for writing the thesis.
See examples of relevant research was from previous thesis.
Possaghi, I., Vesin, B., Zhang, F., Sharma, K., Knudsen, C., Bjørkum, H., & Papavlasopoulou, S. (2025). Integrating multi-modal learning analytics dashboard in K-12 education: insights for enhancing orchestration and teacher decision-making. Smart Learning Environments, 12(1), 53
Zhang, F., Brynildsrud, H., Papavlasopoulou, S., Sharma, K., & Giannakos, M. (2025). Where inquiry-based science learning meets gamification: a design case of Experiverse. Behaviour & Information Technology, 44(5), 1099-1121.
E. Chatzidaki, E. P. Nguyen Doan, E. T. Kjelstrup, S. Papavlasopoulou and M. Giannakos, "Gamification in Informal Science Education: Enhancing Children's Motivation and Engagement with VitenChallenge Application," 2025 IEEE Global Engineering Education Conference (EDUCON), London, United Kingdom, 2025, pp. 1-10.
Dahl Aarhus, K., Motland, J. H., Zhang, F., & Papavlasopoulou, S. (2023). Development and Evaluation of a Gamified Application for Environmental Education: coralQuest. In Interactive Mobile Communication, Technologies and Learning (pp. 374-383).
Open source software (OSS) has become a cornerstone of modern software development, driving innovation, collaboration, transparency and accessibility across industries. However, their sustainability practices often vary widely and there are ongoing challenges to ensure the long-term sustainability of such projects. Sustainability in this context involves community engagement, technical maintenance, governance, policies, and contributor retention, among other issues.
The thesis will explore the key factors that contribute to the sustainability of open source projects and identify actionable strategies that improve their long-term viability. This thesis aims to bring a comprehensive evaluation of the sustainability practices within open source projects, focusing on specific dimensions, e.g., environmental, social, and economic aspects.
Suggested research question for this study could be: What are the key social, technical, and organizational factors that influence the sustainability of open source projects, and how can best practices be developed to ensure their long-term success?
The research begins with an extensive literature review to examine the status for OSS’s sustainability. Then, the thesis will analyze selected successful/ unsuccessful OSS projects to identify patterns in sustainable practices and failure points. Based on qualitative interviews with OSS project maintainers, contributors, and community leaders, data will be gathered to provide insights into effective strategies for sustaining open source projects. Alternatively, surveys could be distributed across OSS communities to collect broader data on contributors' experiences, motivations, and challenges. Expected outcome of the thesis is to provide actionable insights to improve the longevity and resilience of open source projects.
This task is done autumn 2027 by Espen Iversen, other tasks for 2027 are not decided
In Norway we have a well-developed standard for naming equipment and components in buildings, TFM. However, abroad there is no such standard and many different conventions are created and used.
In this project the student will use our properly labeled (ground truth) time series data to build a model that can classify equipment based on the data they've emitted the past two weeks or less. The student will have to settle on a suitable way to preprocess the time series and type of model to use. Also, perhaps evaluate if the classes should be joined or divided based on clustering. This is challenging because different buildings may have different patterns of operation and setpoints.
The student will either get access to our API to fetch data or a hard drive containing the data. He/she will also have a list of labeled tags (sensors). We have years of high frequency data from hundreds of
buildings. The data quality in Building Automation Systems is generally very good. The data will be anonymized but not require any further level of protection.
From simple array iteration to implementing advanced machine learning kernels, loops appear ubiquitously in software. While this is a trivial observation, it has deep performance implications: a considerable amount of execution time in many programs is spent within loops. As a result, optimizing loops has always been a central concern in both compiler design and computer architecture. Traditional software-based approaches rely on techniques such as loop unrolling or loop-invariant code motion. However, in embedded and resource-constrained systems, these optimizations are often insufficient. Many of these systems rely on scalar architectures, where minimizing control overhead becomes particularly important. To tackle this issue, some architectures provide hardware loops, specialized mechanisms that allow loops to execute with minimal branch overhead by managing loop control directly in hardware.
The CV32E40P core [1], developed by the OpenHW Group, is a 32-bit RISC-V core [2] designed for embedded applications. Despite the availability of hardware loop support [3], current compiler toolchains (both GCC and LLVM) provide suboptimal support for exploiting these features. In practice, this results in missed opportunities to use hardware loops that could be generated but are not, or in cases where existing compiler optimizations prevent their formation.
A key challenge lies in the constraints imposed by hardware loops in the CV32E40P architecture. Hardware loops are not a general-purpose replacement for arbitrary software loops, but they require the loop body to satisfy specific conditions. For instance, loops must conform to restrictions on the instructions contained within the loop body. Additionally, only a limited number of nested hardware loops are supported.
These constraints make it challenging for compilers to automatically map arbitrary high-level loops to hardware loops. Furthermore, current toolchains lack a systematic approach to reshape or guide loop transformations in a way that maintains or enhances the applicability of hardware loops. Consequently, there exists a gap between the capabilities of the hardware and the effectiveness of the compiler in utilizing those capabilities.
The goal of this project is to investigate and develop methodologies for efficiently generating hardware loops within an LLVM compiler toolchain [4], specifically targeting the CV32E40P architecture as a use case.
The project will explore:
1. Detection of Loop Patterns: Identifying loop patterns that are suitable for hardware loop conversion.
2. Loop Transformation and Optimization: Transforming and optimizing loops to maximize the applicability of hardware loops.
3. Integration of Hardware-Specific Knowledge: Incorporating a minimal set of hardware-specific knowledge into a target-agnostic compiler infrastructure.
Additionally, the project will examine the role of MLIR (Multi-Level Intermediate Representation) [5] in facilitating more effective loop optimizations. MLIR is a compiler framework that extends LLVM by introducing multiple levels of abstraction for representing programs. This allows for transformations to be applied at higher semantic levels before lowering the representation to machine code. This aspect is especially relevant to the project, as MLIR provides structured representations of loops, such as affine loops with explicit bounds and dependencies. By leveraging MLIR, the project aims to investigate whether higher-level loop representations can enhance the detection and generation of efficient hardware loops compared to using traditional LLVM IR alone.
This project involves deep engagement with international collaborators based in:
Programming Languages: C/C++ (mandatory)Tools: CMake, GitOS: *nixCompiler toolchain: LLVM (desired)English language: working proficiency (mandatory)
[1] CORE-V CV32E40P User Manual (https://docs.openhwgroup.org/projects/cv32e40p-user-manual/)
[2] The RISC-V Instruction Set Manual (https://riscv.org/technical/specifications/)
[3] CORE-V CV32E40P User Manual: Hardware Loops (https://docs.openhwgroup.org/projects/cv32e40p-user-manual/en/cv32e40p_v1.5.0/corev_hw_loop.html)
[4] LLVM Project (https://llvm.org/)
[5] MLIR: A Compiler Infrastructure for the End of Moore’s Law (https://arxiv.org/abs/2002.11054)
Teamwork is central in software development, and is currently a topic much addressed in agile software development where the development is performed in small, self-organized teams. Improving the efficiency and effectiveness in software development will therefore often involve improving the way the teamwork is organized. Several team performance models have been suggested in the research literature, and there is a growing number of empirical studies of teamwork in software development with focus on specific characteristics for agile development teams and distributed or virtual teams.
This assignment will involve making a literature review of research on teamwork in software development, and can be continued in a master thesis with an empirical study of factors that effect team effectiveness in software development teams. For an overview of previous findings on teamwork, see: ieeexplore.ieee.org/abstract/document/7498535
Many companies are developing products to help Human Resources (HR) departments be more efficient in their work. A lot of their time is consumed by reoccurring questions from their employees and managers.
For years HR software have tried to tackle this problem by using manual labor. However, the problem lies in accessibility and the fact that is easier to ask the question directly and get a qualified answer to your problem on the fly, does not make it a sustainable approach.
Meanwhile, GPT (ChatGPT and other Generative Pretrained Transformers) is increasingly able to create consistant answers to general answers in realtime, both in English, Norwegian and other languages.
The current task is to asses the possibility of using GPT also on sensitive data. The project can be to automate our manual assistant (going onto maternity leave now) and make a "Mia-prototype" for actual use at NTNU.
IDEAS:
Lag et grensesnitt for uthenting av snittlønn for enkeltmedlemmer.
Short-term wind power forecasting is essential for operational planning but is challenged by noisy and missing sensor data. In Equinor’s wind operations, wind-meter measurements are currently used to predict power production with a short prediction horizon (≈1 hour), where sensor outages can significantly degrade forecast reliability.
This thesis investigates the work of data governance around developing robust short-term wind power forecasting, with a focus on how the organization is handling wind forecasting in spite of irregular time series and sensor failures. The work aims to improve forecast governance and stability under degraded data conditions, and how this perspective can be paired with reliable ML deployment in energy systems.
The research is empirical. Relevant data collection and analysis methods include qualitative, quantitative, or mixed-method approaches. As a starting point, the work will be based on interviews with domain experts in the offshore wind industry and personnel working with wind/energy management.
The research project will be a collaboration with Equinor.
To achieve the above goals, the thesis includes (specific aims to be defined in dialogue with the student):
i. A state-of-the-art literature review on data governance of real time dataii. Development and implementation of governance practices that can support deep learning models for short-term power prediction under degraded sensor conditions using interviews with industry domain staffiii. Design of governance strategies for organizationally coping with sensor outages and weather measurements
Supervision will be in collaboration with Adjunct Professor Vidar Hepsø (AIT)
The carbon emissions of the ICT sector are on par with the emissions of the aviation industry [1]. In addition, energy is becoming an increasingly scarce resource, but the energy consumption of the ICT sector is growing rapidly. It is currently unclear if this growth can be covered by renewable energy sources, and, if it cannot, how energy should be prioritized between ICT and other energy-consuming industries. There are also significant concerns regarding the water consumption [2] of ICT and its public health impacts [3].
The focus of this project is to examine the sustainability of ICT from a Norwegian perspective. While Norwegian energy is nearly completely carbon-free, computing infrastructure supply chains are highly international. The use of computing in Norway thus might have adverse effects in other parts of the world, and, in terms of global warming, it does not matter where greenhouse gases are emitted.
In this topic, we will explore these issues using recently proposed approaches such as ACT [4] and FOCAL [5]. This project is suitable for one student.
[1] https://www.cell.com/patterns/fulltext/S2666-3899(21)00188-4
[2] https://arxiv.org/abs/2304.03271
[3] https://arxiv.org/abs/2412.06288
[4] https://dl.acm.org/doi/abs/10.1145/3470496.3527408
[5] https://dl.acm.org/doi/abs/10.1145/3620665.3640415
While there has been a long discussion about the potential of using Artificial Intelligence in private organizations, now more and more public organizations are implementing solutions to support their operations. From uses for fraud detection, chatbots, autonomous vehicles, or infrastructure monitoring, AI is gaining ground in applications for public administration. This project will be done in connection with SINTEF Digital and will involve data collection, analysis and reporting. The aim is to find out what is the status of AI adoption, what are the potential interesting uses, and what is the value that is realized.
Today more and more companies are using big data analytics to support or drive their business strategies. Yet, there is ongoing debate about whether such investments do indeed create value if so how this value can be captured. The objective of this master thesis is to perform a quantitative study on companies in Norway and examine the ways in which they are applying big data analytics to create business value. The project is in cooperation with the Big Data Observatory (https://www.observatory.no), during which you will learn how to develop research methods and analyze quantitative data.
Kommunikasjonsferdigheter i IT-utdanning trenes i stor grad gjennom gruppeprosjekter, som er en god læringsmetode for dette, men hvor man kan ha en lang tilbakemeldingssyklus eller være usikker på om kommunikasjonen ga et vellykket resultat. Flervalgsspørsmål (multiple choice) har større potensial for kjapp tilbakemelding og mengdetrening fordi ett svar er riktig og de andre feil, men en utfordring er at hvis den flinkeste studenten i gruppa rett og slett vet svaret på spørsmålet eller kan tenke seg fram til det på egen hånd, er ikke et vellykket resultat avhengig av kvaliteten på kommunikasjon i gruppa, bare forhåndskunnskapen medlemmene kom inn med. Målet i dette prosjektet er å utforske om det går an å lage oppgavetyper i IT-fag som er forholdsvis små og kjappe og løse og hvor man ser om svar ble riktig eller ikke, men hvor man likevel er fullstendig avhengig av god kommunikasjon mellom de involverte studentene for å få riktig løsning, uavhengig av om noen i gruppa har mer kunnskap enn andre. Man kan da lage en prototype for et verktøy (f.eks. læringsverktøy, spill, …) for slik kommunikasjonstrening – og undersøke i hvilken grad studenter føler at det gir verdifull og relevant trening med muligheter for gradvis forbedring. En mulig idé er at den nødvendige informasjonen som trengs for å løse problemet, er fordelt på gruppedeltagerne slik at ingen har den komplette informasjonen – og de heller ikke kan vise hverandre informasjonen, men f.eks. må kommunisere den muntlig – slik at alle må bidra i kommunikasjonen for a gruppa skal komme fram til riktig løsning.
The majority of modern applications are written in the so-called high-level productivity languages such as Python, NodeJS, Javascript, etc. In contrast, computer architecture and hardware research is mostly driven by software written in compiled languages such a C, C++ etc. The mismatch limits our understanding of how these applications are executed on the hardware/processors. For example, while the code written in C, C++ is handled by the “front-end” structures like instruction cache, branch predictors etc. of a processor, Python and NodeJS application code is handled by the “back-end” structures like data cache. This is because Python and NodeJS runtimes/interpreters are treated as code at hardware level, while both the application code and data are treated at data. Consequently our understanding of how to build efficient hardware/processors for the bulk of these applications is limited.
To achieve the level of understanding needed, we require better tools to measure the behaviour of such workloads throughout the computing pipeline. This project is concerned with designing and exploring the space for such a tool that can precisely pinpoint which cache lines contain applications data and which ones have the application code. While the tool helps us track the information in the front end, we must understand its effects on the “front-end” and “back-end” components. Understanding the impact of application behaviour on these components is of utmost importance to address their inefficiencies. Understanding their execution behaviour will allows us to propose new methods that not only ensure the efficient execution of such applications from performance perspective but also with regards to energy-efficiency.
This project aims to explore how sleep-tracking technologies influence individuals’ sleep awareness, behaviors, and well-being by examining the interaction between users and sleep-tracking systems.
Sleep tracking has become a rapidly growing form of self-tracking through wearable devices and mobile applications. While these technologies promise to improve sleep awareness and support healthier sleep habits, their actual impact on users’ behavior and well-being remains unclear. Sleep-tracking tools may empower users through personalized insights, feedback, and data visualizations that promote reflection and behavioral change. At the same time, they may introduce challenges such as over-reliance on metrics, anxiety about sleep performance, or misinterpretation of data. Understanding how individuals engage with sleep-tracking technologies and how these interactions shape outcomes is therefore crucial.
The project consists of a literature review in the area of sleep tracking, self-tracking technologies, and digital health. This will help narrow down the specific research focus, for example exploring individual differences in sleep-tracking adoption, behavioral change mechanisms, or the design of user-centered sleep-tracking tools.
The second phase includes designing an empirical study and collecting data using qualitative and/or quantitative methods. For example, the project may investigate how users interpret sleep data from wearable devices, how sleep-tracking feedback affects daily habits, or how design features influence engagement with sleep-tracking applications. Possible approaches include surveys, experiments, diary studies, or interviews with sleep-tracking users.
In the final phase, the student will analyze the collected data and write up the thesis, contributing insights into how sleep-tracking technologies can be designed and used to better support sleep health and user well-being.
Feng, S., Mäntymäki, M., & Pappas, I. O. (2026). Sleep tracking: an integrative review, conceptual framework and future research agendas. Behaviour & Information Technology, 1-31.
*If you like the topic but you are thinking of a different approach, feel free to send me an email with your suggestions.
Description:A web-based platform that allows users to manage multiple social media accounts (Instagram, Tiktok, Facebook, X, LinkedIn) from one dashboard. Users can create, edit, schedule, publish, and analyze content efficiently.
Objective:To simplify social media management, improve efficiency, and provide centralized tools for content creation, editing, and performance tracking.
Something we humans rarely notice, is how our eyes are in constant movement, and all of these constant movements can be put into four categories. Eye tracking is a technology used to capture an individual's eye movement, and is most commonly achieved by having a small infrared camera for each eye, and then use the center of the pupil as a starting point for further calculations to categorize the eye movements. Unfortunately, this categorization remains expensive, as it has to be conducted by specialists and is time consuming. An easily mistaken assumption in this regard, is that this is a task which would be trivial to automate by defining parameters for what makes a movement fall into a given category and simply use these algorithms to perform the classification. The reality is, however, different and considerably more complex. There are several reason for why this is more complicated that what one might initially assume, but inaccuracies in the captured data (e.g. due to hardware, pupil center algorithms, cornea reflections, etc.) is the most prominent one.
The thesis will be to map already existing, open, and annotated data sets, and to assess these from a quality perspective. Further, a look into how using different types of unsupervised learning, augmentation of the imagery, and other beneficial techniques/means, can benefit the accuracy of automatic annotation. A comparison of the annotations of the already existing and open datasets, to that of the automatic annotations performed, is then to be undertaken.
Additionally, it's desirable to explore the possibility of using the developed methods for eye tracking in VR headsets, as this introduces an extra dimension where the eye tracking cameras themselves will move slightly relative to the eye, due to the head movements which occur when moving about in VR.
This is open to students who have an interest and an idea for a project in Visual Computing.
Students will be chosen based on the following criteria:
Visual SLAM is a term for a set of methods and algorithms that a) determine the motion of a camera (or a set of cameras) through an environment and b) determine the geometrical shape of that environment. vSLAM often builds on detecting “prominent points” in the images, and tracking them through the sequence. If a sufficient number of such points are tracked between two images, the relative pose (=translation and rotation) of the camera can be estimated. As any measurement in images is afflicted by errors, both these pose estimates as well as the estimated 3D positions of the observed image points are uncertain, and the estimation of the complete camera trajectory as well as the scene model “stitched together” from many views needs to be input data to a huge optimization problem.In AROS, we have access to both real video footage from underwater missions, as well as a realistic simulation environment which is able to generate video sequences where the motion and the 3D geometry are precisely known (‘ground truth’). The student project is integrated into our design and development process for a vSLAM system which is specifically tuned to be able with the substantial problems of underwater video material: limited visibility due to turbid water, bad illumination which is also moving with the robot vehicle, disturbances by plankton, dirt, and small fish, and many more. Which part of the vSLAM development is determined to be the focus area of the student project is subject to negotiation; the intention is to let the students experiment with novel approaches proposed in the recent literature, some of them focusing on geometric models and statistical estimation theory, others on machine learning. So we are able to adapt the topic largely to the background knowledge the student(s) already have, and their interest into different relevant research fields, such as e.g. state estimation, optimization, object detection and tracking, machine learning and deep learning.
Potential focus topics:* Robust keypoint tracking in the presence of underwater image degradation* Dynamic Model based prediction and correcting in keypoint and object tracking in underwater conditions* Pose graph and state sequence optimization for underwater visual SLAM* Integration of IMU measurements in underwater visual SLAM* Machine Learning for depth estimation, flow field estimation, and visual clutter detection
Literature:
D. Scaramuzza, F. Fraundorfer: Visual Odometry: Part I - The First 30 Years and Fundamentals. IEEE Robotics and Automation Magazine, 2011.F. Fraundorfer, D. Scaramuzza: Visual odometry: Part II - Matching, robustness, optimization, and applications. IEEE Robotics and Automation Magazine, 2012.
Cesar Cadena, Luca Carlone, et al.: Past, Present, and Future of Simultaneous Localization And Mapping: Towards the Robust-Perception Age. 2016
H. Zhan et al: DF-VO: What Should Be Learnt for Visual Odometry? 2021
Over the last years the emergence of key technologies such as big data analytics and artificial intelligence have given rise to a completely new set of skills that are needed in private and public organizations. With IT gaining an increasingly central part in the shaping of business strategies, it is important that study curricula follow these requirements and provide graduates that fit the needs of organizations. This project will be run in collaboration with the Big Data Observatory (https://www.observatory.no) and involve collecting data through focus groups and surveys with key representatives. The output will involve a detailed look at what skills are necessary and how they can be addressed by educational institutions.
While there has been a lot of focus on the technical aspects related to artificial intelligence, recent years have seen a growing discussion about what the application of AI could be for private and public organizations. The objective of this master thesis project is to examine the readiness of private and public organizations to adopt AI, and the value they have derived from such investments. This project will involve collecting and analyzing data in collaboration with the researchers from the Big Data Observatory (https://www.observatory.no). It is an exciting opportunity to see how organizations are planning to use AI and what steps they need to take to adopt such technologies.
The aim of this work is to design extended reality (AR, VR, MR) tools and applications for serious games and medical applications
Possible topics:
- AI and LLM (serious games that take advantage of LLM models), teaching image processing or computer graphics / deep learning- AI and XR (generating content and assets using AI tools)
- Bronchoscopy related (teaching and planning the procedure)- Echocardiography related (surgical planning)
