Velg hva du ønkser å vise prosjekt for.
Introduction:
[ Vis hele beskrivelsen ]
The emergence of Self-Sovereign Identity (SSI), also known as decentralized technology in healthcare, offers a transformative approach to Electronic Health Records (EHR) management, emphasizing patient autonomy and data security. Similar to the rapid adoption of ChatGPT in artificial intelligence, the integration of SSI in healthcare has gained significant interest, promising a more secure, private, and patient-focused EHR system.
Objective:
This project aims to develop an architecture and prototype for a decentralized Electronic Health Record (EHR) management system based on existing literature. The architecture and prototype will be designed to closely match the current EHR setting in Norway but will operate decentralizedly.
Methodology:
It will have the following components.
These findings are expected to contribute to the existing literature on SSI and Software Engineering. This research will provide knowledge about the state of practice, challenges, and opportunities in integrating tools like decentralized identities into the software development process.
[ Skjul beskrivelse ]
In the evolving landscape of healthcare technology, ensuring patient privacy and security in Electronic Health Records (EHR) is paramount. Self-Sovereign Identity (SSI) offers a promising solution to empower patients with control over their identities and health data.
This proposal aims to contribute valuable insights into leveraging Self-Sovereign Identity for empowering patient privacy and security in Electronic Health Records, ultimately fostering a more patient-centered approach to healthcare data management.
Objectives:
Expected Outcomes:
This project explores how Artificial Intelligence (AI) can enhance game mechanics, player engagement, and social interaction in BitPet, a location-based Augmented Reality (AR) game designed to promote physical activity. The game, inspired by Tamagotchi, Pokémon GO, Animal Crossing, and Pokémon Snap, has been in development since 2020 and is set for a soft launch in Summer 2025.
This project offers an opportunity to work with cutting-edge AI technology while gaining hands-on experience in game development. Developers from BitPet will provide technical support throughout the process.
Project Scope:
The front end will be developed in Unity. This project is designed for two students, and experience with Unity development is beneficial.
This project aims to investigate the physical and social health effects of playing BitPet, a location-based AR game designed to promote fun, physical activity, and social interaction. Inspired by Tamagotchi, Pokémon GO, Animal Crossing, and Pokémon Snap, BitPet has been in development since 2020 and is scheduled for a soft launch in Summer 2025.
Project Phases:
Literature Review – Study existing theories, games, and empirical research on the health effects of similar games.
Empirical Study Design – Develop a study to evaluate:
Data Collection & Analysis – Conduct questionnaires, interviews, observations, and analyze in-game data tracking.
Findings & Recommendations – Provide:
The project will be co-supervised by faculty from NTNU’s Faculty of Medicine and Health Sciences and is suited for one or two students.
This project aims to develop game mechanics that will motivate users to socialize and be physically active using Augmented Reality. It is part of the BitPet project, which aims for commercialization. Developers in BitPet will provide technical support.
The project will involve a study of existing theory, game concepts, and technology, the design and development of a game concept (both front-end and back-end), and an evaluation of the concept involving real users.
The front end will be developed in Unity.
This project requires two students.
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.
Expected Contributions
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).
Real-time object detection and semantic segmentation require models that are both highly accurate and computationally efficient. RTMDet [1], a recent object detection model, strikes a balance between speed and precision, but it can still be computationally demanding when deployed on edge devices.
Efficient deployment of RTMDet on the NVIDIA AGX Orin platform presents several challenges, including memory bandwidth limitations, power efficiency constraints, and real-time processing requirements, especially when handling high-resolution video feeds from a multi-camera system.
While Orin integrates GPU and DLA cores, the optimal utilisation of these resources demands careful tuning of computation pipelines.
This thesis aims to develop optimisation techniques to enhance the performance of RTMDet on the NVIDIA AGX Orin, ensuring real-time performance with minimal computational overhead.
- Analyse the computational bottlenecks of RTMDet on NVIDIA Orin.
- Explore various optimisation techniques, such as model quantisation, pruning, tensor decomposition, and GPU-DLA offloading.
- Implement and benchmark RTMDet on Orin, comparing performance in terms of speed, power consumption, and accuracy.
- Integrate the optimised RTMDet model into the SeaSIght pipeline.
Literature Review: Study existing research on RTMDet and optimisation techniques.
Baseline Implementation: Implement and integrate RTMDet with SeaSIght pipeline on NVIDIA Orin using TensorRT and CUDA, and analyse performance metrics.
Optimisation Techniques: Apply and evaluate various optimisation strategies
- TensorRT Optimisation: Utilise FP16 and INT8 quantisation to improve speed and reduce memory footprint.
- Model Pruning: Remove redundant parameters to accelerate computation while maintaining accuracy.
- GPU-DLA Offloading: Partition model layers between the GPU and DLA to optimise parallel execution or convert the model to run fully on DLA cores.
- Memory Optimisation: Use efficient memory access patterns and kernel fusion to reduce latency.
- Profile execution: Use nvprof and Nsight Systems to identify bottlenecks.
Benchmarking and Evaluation: Compare the optimised implementations against the baseline in terms of inference latency, throughput (FPS), resource utilisation, energy consumption, and accuracy. Document findings and potential further optimizations.
Hybrid Optimisation Framework: Develop a strategy that combines the most effective techniques to achieve optimal real-time performance on NVIDIA AGX Orin and seamlessly integrate it into Maritime Robotics’ SeaSight pipeline.
1. An optimised implementation of RTMDet achieving real-time inference with minimal energy consumption on NVIDIA Orin.
2. A set of optimisation techniques tailored to RTMDet for efficient deployment on Orin’s GPU and DLA cores.
3. Integration of the optimised RTMDet implementation into the SeaSight system, enhancing situational awareness and providing real-time insights to USVs operators.
Hardware: Maritime Robotics’s SeaSight module, NVIDIA Jetson AGX Orin/ Jetson Orin NX.
Datasets: custom-made dataset (or COCO, Pascal VOC for object detection benchmarking).
[1] https://arxiv.org/abs/2212.07784
https://github.com/open-mmlab/mmdetection/tree/3.x/configs/rtmdet
This project aims to design and develop innovative game concepts that integrate an exercise bike as a game controller, complementing traditional button inputs. In addition to button controls, players should use pedaling as a core mechanic to interact with the game. The goal is to create an engaging experience that remains enjoyable over time while promoting physical activity.
The game will be developed in Unity.
This project is designed for a team of two students, requiring prior experience with Unity.
The goal of this project is to design and develop new game concepts for a game where an exercise bike is used as a game controller in addition to traditional game input through multiple buttons. In addition to input from buttons, the player should control the game by using her/his fit to move the pedals. The goal of the game is to have fun that can last over time and get physical exercise. The game should be implemented in Unity using an API provided for the exercise bike controller.
The goals of this project are:
Research existing exergames and games that could fit this purposeDesign and implement a prototype game Provide input on the API used for the exergame frameworkEvaluate the game through user experiments
This project is for a group of two students, and experience with Unity is required.
In this project, the goal is to develop new game concepts and technologies for exergames - games where the player performs physical exercise. There are several approaches to exergames, and the challenge is to find the balance between something that is fun to play and getting real physical exercise from playing the game.
The first phase of the project will consist of a theoretical study of exergames and mechanisms for using games as motivators. The second phase focuses on implementing a prototype using various technologies. In the third and final phase, the prototype will be evaluated and tested.
This project requires a group of two students.
This project aims to develop innovative game concepts and technologies for exergames—games that incorporate physical exercise as a core mechanic. A key challenge is balancing engaging gameplay with meaningful physical activity to create a fun and effective experience.
The project will be carried out in three phases:
This project requires a team of two students.
This project aims to design, implement, and evaluate a multi-player learning game where students work together or compete to complete challenges while simultaneously acquiring knowledge. The game must strike a balance between engagement and education, ensuring it remains both enjoyable and effective as a learning tool. The students are encouraged to use AI a key component to improve the gameplay and the experience.
The project will involve reviewing research on game-based learning, games, and use of AI in games, developing and implementing a game concept, and conducting user evaluations to assess its effectiveness.
This project is designed for a team of two students.
Graph domain adaptation (GDA) is an effective approach for knowledge transfer between graph data structures. Typically, in a GDA task, there are source graph(s), and the aim is to adapt knowledge learned from these source graph(s) to target graph(s). It is an emerging area because training graphs for individual tasks is expensive and suffers from label scarcity. The methods for GDA can be categorized into source-based, adaptation-based, and target-based approaches. In this project, we aim to focus on adaptation-based methods, although other methods are also welcome. Adaptation-based methods usually consider shifting between domains such as structural shift, marginal shift, and task shift.
This thesis aims to investigate adaptation shifting mechanisms in Graph Domain Adaptation tasks, with a particular focus on adaptation-based methods. First, the project will identify and characterize different types of domain shifts in graph-structured data. Then, it will evaluate the effectiveness of existing adaptation-based GDA methods in handling these shifts. The project may also explore the development of metrics and/or visualization techniques to measure these shifts. If feasible, the study may also propose novel approaches to mitigate negative effects in this area.
· Literature review: Comprehensive review of existing GDA methods with a focus on adaptation-based approaches and domain shift characterization (or any other method if chosen differently).
· Dataset selection: Datasets will be determined together with the supervisor on a publicly available dataset.
· Implementation: Implementation of existing (state-of-the-art) methods.
· Metric development: Develop techniques to visualize domain shifts and adaptation processes in graph embeddings to enhance interpretability.
· Algorithm development: Develop novel adaptation mechanisms specifically targeting identified shift types.
· Experimentation and evaluation: Conduct a comparative analysis of adaptation methods, including standard metrics as well as domain-specific measures. Novel approaches will also be evaluated if suggested.
· A comprehensive analysis of adaptation shifts in graph domain adaptation.
· New metrics and/or visualization tools to measure domain shifts in GDA tasks.
· Experimental insights on how different shifts affect adaptation-based methods.
· Guidelines for practitioners on selecting appropriate GDA methods based on the characteristics of their source and target graph domains.
· Recommendations for designing more efficient GDA models.
Co-advisor: PhD Candidate Ahmet Tüzen - ahmet.tuzen@ntnu.no
Graph meta-learning is an emerging approach that addresses the challenge of learning from limited labeled graph data. Traditional graph learning methods typically require substantial amounts of labeled data to achieve good performance, which is often impractical in real-world scenarios due to the high cost of data annotation. Meta-learning, also known as "learning to learn," offers a promising solution by leveraging knowledge from related tasks to quickly adapt to new tasks with minimal data. In graph-structured data, meta-learning techniques aim to extract transferable knowledge across different graph learning tasks, enabling models to generalize effectively to unseen tasks with only a few examples. This project focuses on investigating meta-learning strategies designed for graph neural networks (GNNs), exploring how to efficiently capture and transfer structural and semantic patterns across graph datasets.
This thesis aims to investigate meta-learning approaches for graph-structured data, with a particular focus on few-shot learning in graph neural networks. This study will examine how meta-learning frameworks can be effectively adapted to graph neural networks to improve generalization with limited labeled data.
· Literature review: Comprehensive review of existing meta-learning paradigms and their applications to graph-structured data.
· Framework analysis: Systematic analysis of how different graph properties (structure, node features, edge attributes) affect meta-learning performance.
· Implementation: Implementing state-of-the-art graph meta-learning methods as baselines.
· Algorithm development: Design and develop novel meta-learning approaches that specifically address the unique challenges of graph-structured data.
· Visualization: Develop techniques to visualize the extraction and transfer of meta-knowledge in graph embeddings.
· Experimentation and evaluation: Conduct a comparative analysis of different meta-learning strategies across various graph learning tasks (e.g., node classification, graph classification, link prediction).
· A comprehensive analysis of meta-learning approaches for graph-structured data.
· Insights into the effectiveness of different meta-learning strategies for various graph learning tasks.
· Novel algorithms or adaptations of existing meta-learning frameworks, specifically tailored for graph neural networks.
· Visualization tools to interpret meta-knowledge transfer in graph domains.
· Guidelines for applying meta-learning techniques to real-world graph learning problems with scarce labeled data.
Thesis Description:
This master's thesis addresses the challenging problem of learning from limited labeled data in graph domains through meta-learning techniques. By systematically investigating how meta-learning principles can be applied to graph neural networks, this research aims to enhance our understanding of knowledge transfer in graph-structured data. The work will contribute novel insights into designing meta-learning algorithms that can efficiently capture and leverage structural patterns across different graph tasks, which may contribute to more data-efficient learning systems for real-world graph applications.
Description in which company/unit the thesis will be placed: This master thesis will be carried out under the project "Generation of Large-Scale Norwegian GPT-2 Language Models" supported by the Norwegian Research Center for AI Innovation (NorwAI).
Problem Description: Today there are two official forms of written Norwegian, Bokmål and Nynorsk, each with its own variants. Even though the two written languages are considered equal, only 5-10% of Norwegian texts are written in Nynorsk. Nynorsk is therefore a low-resource language within a low-resource language. Translation between Bokmål and Nynorsk is not available in Google Translate, and all this research is very much in its infancy. On the technical side, almost existing neural machine translation (NMT) systems are based on the encoder-decoder framework and require a large sentence-aligned parallel corpus for model training. However, the lack of such data in Bokmål and Nynorsk renders the most commonly used NMT approaches ineffective. Meanwhile, Generative Pre-trained Transformer (GPT), which captures the rich contextual knowledge from large scale web texts, has gained tremendous success in a variety of natural language processing tasks. Thus, the upcoming question is whether the GPT language model can be used for machine translation with only limited parallel data.
Thesis Description: The overall goal of this project and master thesis is to develop an efficient GPT based neural machine translation model for Bokmål and Nynorsk. To achieve this goal, the project is divided into the following tasks: (1) Investigate the possible ways of utilizing pre-trained language models for translation tasks. As part of this, the candidates are expected to perform a state-of-the-art literature review and summarize the strength and weaknesses of existing approaches. (2) Based on the investigation results from (1), the candidates need to design an efficient approach to incorporating the GPT language model into neural machine translation model. This includes evaluating the method with respect to applicability.
Data Description: The Bokmål and Nynorsk parallel data come from the Norwegian Colossal Corpus at the National Library of Norway. This dataset has 200k translation units (TU) i.e. aligned pairs of sentences in the respective languages extracted from textbooks of various subjects and newspapers, giving a total size of 28.6 MB. It is publicly available with the license CC BY 4.0. The GPT language model is provided by the Norwegian Research Center for AI Innovation, which is trained from scratch for Norwegian and has 318.6M parameters.
Challenges (business and/or research): (1) With a small bilingual training corpus, it would be hard for the NMT model to learn how to recognize and translate named entities, which might sacrifice translation performance. (2) Most of the instances in the training corpus are short texts. How to deal with the translation of long texts is another challenge. (3) Fine-tuning is the prevalent paradigm for using large pre-trained language models to perform downstream tasks (e.g., translation and summarization), but it requires updating and storing all the parameters of the LM. Considering the size of current LMs, how to reduce training time and storage overhead when building and deploying machine translation systems is an urgent problem to be solved.
Supervisor (NTNU): Lemei Zhang (lemei.zhang@ntnu.no), Peng Liu (peng.liu@ntnu.no)
The goal of this thesis is to explore the use of symbolic AI (e.g., case-based reasoning) to enhance the reasoning capabilities of GPT to reduce hallucinations and opaqueness, and improve trustworthiness.
The latest versions of GPT are achieving unprecedented capabilities for assisting humans in completing diverse and complex tasks. Despite their extraordinary performance, GPT models remain opaque and prone to hallucinations. Moreover, GPT models are still underperforming when completing complex reasoning processes. These problems are due to the lack of factual knowledge because, during training, GPT memorizes facts defined in the training data, but it is not always able to recall the correct facts and often experiences hallucinations by generating factually incorrect statements [Pan 2024]. This issue severely affects users' confidence in applying GPT for complex tasks. To handle these shortcomings, the research community and the industry are exploring the possibility of integrating symbolic AI methods, which include rule and reasoning engines and knowledge graphs. There is a consensus that symbolic systems and Large Language Models (LLMs), such as GPT, can be considered complementary to each other’s: while the former ones are human-interpretable, deterministic, and parsimonious in terms of data, the latter ones are inherently opaque, indeterministic, and data-hungry. At the same time, while symbolic approaches often require human experts to manually encode symbolic knowledge, LLMs, such as GPT, typically support some form of automatic learning from data [Calegari 2020].
The goal of this thesis is to explore the use of symbolic AI (e.g., case-based reasoning or other reasoning engines) to enhance the reasoning capabilities of LLMs, such as GPT, reduce hallucinations and opaqueness, and improve trustworthiness. At the same time, exploring how LLMs can augment symbolic systems by providing natural language interfaces and reducing the cost of building that support these systems would be interesting (aka knowledge). This might require applying advanced prompt engineering techniques and fine-tuning.
After completing the MSc thesis, the student should learn how to improve the performance of GPT using prompt engineering techniques, fine-tuning, and symbolic systems such as CBR systems.
The expected results of the thesis include the implementation of a prototype that integrates a GPT model with a symbolic AI system. This might include, for instance (non-exhaustive list):
In addition, the student must validate the prototype's performance using existing benchmarks or create new ones tailored to the specific task to be solved.
Candidates should have a good understanding of deep learning techniques, data engineering, and case-based reasoning. Moreover, it is recommended to have experience in programming in Python with libraries for data processing (e.g., Pandas, SQLAlchemy, etc.), data analytics (NumPy, Scikit-learn, TensorFlow, PyTorch, etc.), and data visualization (e.g., Matplotlib, Seaborne, etc.).
Some relevant courses at NTNU: TDT4171, TDT4173, TDT55
This project will be a collaboration between NorwAI and Sintef. Supervisors are Kerstin Bach, Francisco Martin-Recuerda, and Erik Johan Nystad. If you have questions, please get in touch with us.
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.
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.
This thesis focuses on classifying an object through minimal tactile exploration. That is to infer the shape or class of an object based on a sequence of touches by a tactile sensor. The goal is to develop a next best touch (NBT) strategy that guides the exploration process by selecting the most informative touch points to maximize shape inference accuracy. Instead of random exploration, the system will leverage learned priors to make data-driven decisions about the optimal place to touch next.
The classification task involves distinguishing between a fixed set of objects, where the system must infer which object is being touched/grasped after a fixed number of touches (e.g., five). A cost-function-based approach will be implemented to optimize the NBT strategy, with the potential integration of learned heuristics (pure learning-based approaches are also acceptable) to refine the decision-making process. The final model should be capable of efficiently reconstructing object geometry and making accurate classifications with minimal interaction.
This research has applications in robotic tactile perception, particularly in scenarios where visual sensing is limited or unreliable, such as identifying objects in cluttered or occluded environments. The MSc thesis entails signing an agreement with SINTEF Ocean.
Proposed approach
The proposed approach involves collecting tactile data using a GelSight sensor mounted on a robotic arm while interacting with various objects. This dataset will be used to train a classifier model capable of inferring object identity based on a sequence of sensor readings. The system will then learn to identify the most informative next touch (NBT) that maximizes classification accuracy.
For validation, the trained model will be tested on a fixed set of relevant objects with different shapes, where the system must correctly identify which object is being grasped using a limited number of tactile interactions. The classifier’s performance will be evaluated based on its efficiency in making accurate inferences with minimal touches.
Requirements
Supervisors: Theoharis Theoharis (NTNU), Ekrem Misimi, Sverre Herland (SINTEF Ocean)
Reference
https://ieeexplore.ieee.org/document/10801324
Interested in the future of real-time photorealistic reconstruction of complex 3D scenes?
In that case this is an excellent opportunity to understand, experiment with and explore the technology, for example in the context of autonomous driving and simulators.
Some references:
More info here
In a digital age, where users are faced with a significant amount of data, the recommendation system has become an essential part of everyday life to assist users in selecting products and services that are suitable for their needs. The music industry, in particular, has seen a rapid growth in the recent years, with users now able to access a greater variety of content. The need for user-tailored recommendations is, therefore, a necessity.
Students can choose their cio-inspired AI technique of choice and their path within music recommendation or other recommendation systems.
A couple of ongoing/ completed relevant projects
More healthcare applications are using AI. Many of these applications are categorized as high-risk. Thus, it is essential to educate stakeholders in the healthcare domain to understand the EU AI Act. However, the EU AI Act is a complex regulation that is hard to follow. This project aims to study how to design and develop an educational game to teach the EU AI Act. The project will apply the design science research method and invite healthcare stakeholders to pilot and evaluate the game. The expected results are the prototype of the game and the methodology to design such a game.
The project will be co-supervised by Prof. Øystein Nytrø and Prof. Alf Inge Wang.
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.
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.
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
Training of AI agents in virtual environments using deep reinforcement learning.
Some examples of virtual environments that we are using: CARLA, NVIDIA Omniverse (e.g. Isaac Sim or DRIVE Sim) and MARS, etc.
Some references to SotA approches that we are looking into:
Additional industri references:
We are also working to create a realistic virtual replica / twin of the area around the campus (Gløs) to use in simulators like CARLA and DRIVE Sim for training of AI agents in a local environment.
Various aspects related to AVs, including the use of NAP-lab's new research platform for AD.
End-to-end approaches to AD like imitation and reinforcement learning.
Modular approaches to AVs (i.e. mapping and localization, perception and prediction, planning and control), including the use of the functonality provided by the NVIDIA DRIVE software stack and DRIVE / ROS integration.
The use of simulated environments (like CARLA and NVIDIA DRIVE Sim), new ways to generate a realistic virtual environment using real world data (e.g. NeRF), and domain adaptation from simulated to real world scenarios.
AVs as mobile sensor platforms (assessing the condition of road objects to update Digital Road Twins) and AD in a challenging nordic winter environment (both in collaboration with SVV)
Point Cloud processing (real-time detection and tracking of objects like pedestrians and vehicles in LiDAR data, generation of point-cloud references (e.g. SLAM), and updating this reference as well as re-localize relative to it, collaboration with karverket).
Remote control of an AV using 5G (e.g. management of an AV fleet)
+++
If this is a topic of high interest to you the most effective way would be to have a talk and see if we can find a project highly motivating for you at the same time as the overall NAP-lab project moves forward.
Main investigators for the NAP-lab project is Frank Lindseth and Gabriel Kiss
Additional info
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:
Agentic AI utilizes LLMs and other agents to complete complex tasks. Compared to LLMs, Agentic AI exposes more attack surfaces and is vulnerable to more complicated attacks (e.g., indirect prompt injection attacks). However, the studies on the security of Agentic AI are lacking and immature.
This project aims to systemize the Agentic AI security knowledge, develop approaches to identify the Agentic AI vulnerabilities, and propose countermeasures to mitigate the risks of Agentic AI.
The research methods will include a systematic literature review and design science. The project's expected outputs include approaches and prototypes to identify and mitigate Agentic AI risks.
The project requires good coding skills and AI and Cybersecurity knowledge.
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.
Many organizations struggle with the right way to do digital transformation. "Big bang" methods are often costly and bear high risks of failure too late in the process. Ideas from agile are gradually entering into organizational digital transformation as an alternative to big bang approaches.
In this task we are interested in learning more about what agile transformation means, and how it is practiced successfully by organizations. You will be expected to analyze existing research in agile transformation and digital transformation using qualitative literature review methods. Later on we will together choose a real-world case organization for a field study involving interviews and observations. You will generate new knowledge, and recommendations for how agile information should and should not be used by organizations who want to implement digital transformation.
This task requires that you have a good understanding of, and are interested, in empirical qualitative research. Working language for this task is Norwegian or English. The thesis can be written in Norwegian or English but we recommend English. Please contact Babak before you select this task.
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.
To research and develop an AI tool that can assist teacher in course planning, assessment and evaluating students learning outcomes given course description. AI as a co-teacher has been gaining popularity with generative AI and the availability of readily available pre-trained LLMs. These tools if used ethically could help and assist teachers in improving their tasks; be it course planning, content creation, assessment or evaluation. This project aims to test and deploy domain specific LLM for improving curricula and course connect with respect to learning objectives and to assist teachers in their daily routines.
Prior knowledge: Machine Learning, Generative AI, LLMs
Skills required: Python programming
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.
This project works with diagnostics of eyesight issues and spectacles. It is carried out in close copperation with an international start up working with new ways of addressing eyesight problems.
If you are interessted in computer vision, eyesight, industrial applications and impact in the 3rd world, this is a project for you.
Please contact the supervisor before applying.
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.
The rapid integration of artificial intelligence (AI) into everyday life is reshaping social practices and technological infrastructures. Women remain underrepresented in the development and use of AI, enabling systems to reproduce existing gender biases. Yet AI also holds potential for empowerment by strengthening digital confidence, competence, and agency. Investigating how AI can foster more inclusive and equitable technological futures is therefore essential.
Several projects and master theses have developed interventions for Women in Computer science and AI specifically.
This master's thesis will build on an intervention that explores how an AI hackathon can be intentionally designed to empower women in AI. The focus in this intervention must be on women who do not necessarily study computer science or work in IT with degrees or experience in any subject area. The results will be a set of actionable design principles for developing a hackathon and prototypes (developed by the hackathon participants) that will demonstrate those principles in practice.
In this master thesis, the hackathon will be evaluated and re-engineered.
The research question is: How can a hackathon be designed and implemented to facilitate the empowerment of women in AI?
In the preliminary phase, the student will run a rapid literature review to examine existing research on the intersection AI, hackathon, women, empowerment. Then, the study will investigate technological solutions to build the hackathon. Technology solution ideas and hackathon principles will be generated through rounds of participatory co-design workshops. The master student will have to recruit participants to the hackathon and run the hackathon in cooperation with the SBS group.
Relevant sources:
O. T. Aduragba, J. Yu, A. I. Cristea, M. Hardey and S. Black, "Digital Inclusion in Nothern England: Training Women from Underrepresented Communities in Tech: A Data Analytics Case Study," 2020 15th International Conference on Computer Science & Education (ICCSE), Delft, Netherlands, 2020, pp. 162-168, doi: 10.1109/ICCSE49874.2020.9201693.
Paganini, Lavinia, and Kiev Gama. "Female participation in hackathons: A case study about gender issues in application development marathons." IEEE Revista Iberoamericana de Tecnologias del Aprendizaje 15.4 (2020): 326-335.
Gama, Kiev, et al. "Hackathons as inclusive spaces for prototyping software in open social innovation with NGOs." 2023 IEEE/ACM 45th International Conference on Software Engineering: Software Engineering in Society (ICSE-SEIS). IEEE, 2023.
R. Prado, W. Mendes, K. S. Gama and G. Pinto, "How Trans-Inclusive Are Hackathons?," in IEEE Software, vol. 38, no. 2, pp. 26-31, March-April 2021, doi: 10.1109/MS.2020.3044205.
Paganini, Lavínia, et al. "Opportunities and constraints of women-focused online hackathons." 2023 IEEE/ACM 4th Workshop on Gender Equity, Diversity, and Inclusion in Software Engineering (GEICSE). IEEE, 2023.
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.
The focus of the thesis is to improve and test an existing intelligent feedback system that helps the students while they are programming. This help should be provided in real-time using the eye-tracking data from the student and the log data from the IDE that the student is using. The challenge is to develop a system that is both effective and efficient in helping the students when they are facing difficulties in programming medium-size software.
Thesis DescriptionIn a first step, the student(s) will design and implement the gaze-aware feedback tool. Afterwards, they will conduct a small user study in order to test the usability of the system with a small 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 basic machine learning and system design. Solid programming skills and an interest in hands-on development and experimentation is also a requirement.Programming skills: Python.
The intersection of artificial intelligence (AI) and healthcare presents an opportunity to enhance medical diagnostics, improve patient outcomes, and streamline the work of healthcare professionals. This master project proposal invites students to contribute to this transformative field by developing an innovative web application. This application will leverage AI technologies to analyze medical data, offering insights ranging from data visualizations to complex diagnostics. The project integrates three key areas: scintigraphy image analysis, blood analysis data, and anamnesis analysis. However, students can choose the direction of their studies and focus on the area(s) of their interest.
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:
Recommended technical skills:
Expected Project Work Packages
All proposed projects are available as both specialization and master’s thesis options. They can be adapted to suit students working individually or in pairs.
MIA (Monitorering-Innsikt-Aksjoner) - Dine data. Din digitale tvilling. Dinhelse
Description
Video
RBK - Sport/Fotbal Analytics
MIC - Medical Image Computing / Analytics (several projects):
Pathology - Breast and Prostate (St. Olav):
AAA - Abdominal Aortic Aneurysms
Myocardial Infarction - AI-Tool for improved managing of patients with myocardial infarction
Myocardial Infarction
An overview of MIC-related items can be found here
Kartverket (TBA)
SVV (TBA)
NINA - Monitoring Norwegian nature loss with satellite-based earth observation and AI
Schibsted - Floorplans
RBK (fotball analytics):
Video presentation
Presentation with sound-clip embedded in each slide (download and listen)
St. Olav / NTNU Med. fak. (medical image analysis):
Abdominal Aortic Aneurysms (AAA): Is Minimal invasive or open surgery best for a given patient. (video)
Abdominal Aortic Aneurysms (AAA): Is Minimal invasive or open surgery best for a given patient. (text)
Brain segmentation from high-res MR images
Kartverket:
Video (for the three projects below)
Bruk av kunstig intelligens til klassifisering av laser punktsky fra flybåren datafangst
Beregning av DOP-verdier for å predikere GNSS målekvalitet
Strømming av punktsky fra database til web viewer
NINA:
Monitoring Norwegian nature loss with satellite-based earth observation and AI (Video)
Monitoring Norwegian nature loss with satellite-based earth observation and AI (Text)
MIA Health:
MIA (Monitorering-Innsikt-Aksjoner): Dine data. Din digitale tvilling. Din helse (video)
MIA (Monitorering-Innsikt-Aksjoner): Dine data. Din digitale tvilling. Din helse (tekst)
Catchwise:
Catchwise: Predicting where to find fish for commercial fishing boats
Maritime Robotics:
Dense Monocular Depth Estimation for Unmanned Surface Vessels
The National Archives manages a vast amount of historical data. Some data have limited digital accessibility, such as:
The candidate will explore the use of AI technology, such as Large Language Models (LLMs) and image recognition, to improve accessibility of a selected data source. The National Archives will provide guidance and access to data. The concrete task will be defined after conversation with experts from the National Archives who will also help with the evaluation. Inga Lang (Arkivverket) will support the project as co-supervisor.
Oslo Municipality regularly conducts public consultations (høringsprosesser) to gather feedback on regulations, urban planning, and policy proposals. The Planning and Building Agency (Plan- og bygningsetaten) and other municipal entities receive large volumes of textual input from various stakeholders, including citizens, organizations, and businesses.
Today, processing these responses is largely manual, requiring significant effort to categorize feedback, identify key themes, and summarize insights for decision-making. This process can be time-consuming, prone to bias, and challenging to scale effectively. We hypothesize that AI-powered text analysis can automate and streamline this workflow.
This thesis will explore how AI-based text analysis can improve the processing of public consultation responses. The candidate will work with the Planning and Building Agency (or another relevant municipal agency) to understand the current workflow, identify challenges, and develop a proof-of-concept AI pipeline that automates categorization, clustering of similar feedback, and summarization.
The research will examine the difference in quality between human analysis and the AI pipeline, and assess the changes in processing speed. Oslo Municipality will provide a data set that has the initial responses and the target information allowing the candidate to fine-tune Large Language Models (LLMs).
Pål Enger (Oslo Kommune) will co-supervise the project.
This thesis uses the combination of AI and biometrics (eye-tracking, EEG, Facial expression) to understand processes underlying successful extreme programming (pair programming, test-driven development, continuous integration, refactoring) scenarios. This understanding can help us develop innovative solutions for the
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.
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.
An immune system approach to fake news classification is currently under development . It is an exciting new approach to Fake News classification, drawing inspiration from antibody and antigen concepts from nature. This project seeks to extend and refine the current approach in various ways. The student(s) may choose their own path.
Ongoing relevant projects include:
Description in which company/unit the thesis will be placed: Aneo is a renewable energy company with near 300 employees and headquarters in Trondheim. Aneo aims to produce more renewable energy and develop new, unique products and services to promote further electrification and efficient utilization of renewable energy. AI plays a key role in achieving this goal due to the increasing need for automation and data- driven decision making. We have extensive experience supervising students with weekly meetings and access to AI and domain experts.
Problem Description: Aneo owns and operates 11 wind farms with 225 wind turbines. An important task inoperating wind parks is to forecast how much energy will be produced hourly the next day, socalled day-ahead forecast. This forecast is used to balance electricity production with consumption to maintain stable power grid frequency. The accuracy of the forecast directly influences the profitability of the electricity producer.
Thesis Description: Aneo uses ML to forecast wind power, with weather forecasts as the main input. Weatherforecasts contain multiple weather scenarios, so-called ensemble. Using multiple scenarios as inputs rather than one (the most probable) might improve theaccuracy of the wind power forecast and estimate uncertainty. We want students to find a good way to utilize weather ensembles for probabalistic windpower forecasting.
Project tasks include: - Data preparation and exploration - Building models to forecast wind power production - Conducting experiments to evaluate and compare models - Analyzing experiment results
Description in which company/unit the thesis will be placed: Aneo is a renewable energy company with near 300 employees and headquarters inTrondheim. Aneo aims to produce more renewable energy and develop new, unique products and services to promote further electrification and efficient utilization of renewable energy. AI plays a key role in achieving this goal due to the increasing need for automation and data- driven decision making. We have extensive experience supervising students with weekly meetings and access to AI and domain experts.
Problem Description: Aneo Retail optimizes energy consumption and service costs for cooling systems in retailstores. Part of our work is to detect problems with the equipment in the retail stores and schedulereparation and maintenance. To make this process more efficient we want to to predict faults in advance. This would allow us to optimize maintenance scheduling, order parts and schedule work in away that minimizes consequences.
Project tasks:
• Data preparation and analysis. • Building models to predict future faults, time until faults. • Conducting experiments to evaluate models. • Analyzing experiment results.
Data Description: We will provide two types of data:
We have hundreds of fault events and millions of measurements. All data will be provided in CSV format. The data should be kept in private. We provide cloud infrastructure necessary for the project.
Description in which company/unit the thesis will be placed: Aneo is a power producer with both hydro and wind plants – now we will do more even on wind. Aneo is already a major player in renewable energy for a better society. As the owner and operator of the power plants, Aneo takes care of renewable energy to monitor, control,and optimize the performance of the generation or transmission system or added values through system of computer-aided tools. Among which nowadays AI plays an even more important roles in such system in order to make better decision under increased complexity and risk. AI team in Aneo is responsible for providing the innovated, reliable and robust AI energy system for both Aneo and the third parties, and coordinates Aneo’s AI activities and membership in Norwegian Open AI Lab (NAIL) and participates in research work of the recently established Centre for Research-based Innovation NorwAI.
Problem Description: Operating wind turbines in remote locations is challenging. Aneo experienced bushings,bearings and gearboxes in some wind turbines fail earlier than expected. Such failures can be detected through trend analysis of vibration data and/or other types of data, but the identification of the root cause is a very manual process. The objective of the project is to automate the identification of the root cause to support maintenance decision making for operators. The faults happen frequently enough that the farm owner will benefit significantly from an automated analysis process, but there may not be enough cases for supervised learning to be relevant currently. However, all existing failures have expert description which can help the students better understand the problems. The students are encouraged to use any available technologies to automate the process.
Thesis Description: The project aiming to solving the following tasks:
Data Description: The project is based on existing vibration data from wind turbines. These consist of short periods of high frequency data as well as statistics. Some faults have already been identified in the data. If needed, SCADA data from wind turbines are also available. The data needs to be kept internally.
Comment:
The following description is from The National Audit Office of Norway/Riksrevisjonen. They propose an interesting and important use-case for anomaly detection and want to have the developed system use their real data. At the time of publishing this proposal, it is not 100% guaranteed that the data from Riksrevisjonen that will be made available to us is of sufficient volume and quality for the proposal to be meaningful. Worst case, the data is somehow not meeting our standards, and in that case we will re-focus the project to with other real-life data sources instead (e.g, industrial data from ANEO, transaction data from SMN1). Also, if you are interested in this project, please make sure to look at this page.
Text from The National Audit Office of Norway:
The National Audit Office of Norway (Riksrevisjonen) is the watchdog of the Norwegian parliament, controlling the accounts and efficiency of the government. Our work secures transparency of public spending and helps secure citizens trust in government. This is crucial in a world where democratic values are under pressure.
For students interested in ML-based anomaly detection and financial fraud, Riksrevisjonen offers a unique opportunity to apply data-driven approaches in a real-world context—where insights can directly influence accountability and improve governance.
For more information see: www.riksrevisjonen.no
Name: chief data scientist Jan Roar Beckstrøm
Email: jrb@riksrevisjonen.no
See also: https://www.linkedin.com/in/jan-roar-beckstr%C3%B8m-a8b9901/
Background: The National Audit Office of Norway (Riksrevisjonen) analyzes large amounts of financial data of central government agencies as part of the audit process. Auditors in Riksrevisjonen have started to investigate unsupervised anomaly detection methods, such as isolation forest and autoencoder neural networks, to find unusual transactions in accounting data. Related research[1] shows promising performance of such algorithms in flagging faulty or fraudulent transactions. However, examples of practical application in audits are few.
This thesis should aim at connecting anomaly detection algorithms with audit standards by investigating the following possible research questions:
Accounting data on transactions from Norwegian government entities. It should be possible to provide the same access to accounting data as for the auditors. The data is provided be the government accountant DFØ [2](Direktoratet for økonomistyring), or possible by SIKT[3] if universities/colleges are the target entities.
The data will be raw data, but only limited preprocessing should be necessary. The data is not public data, but contains limited personal data. The student might need store/analyse the data on a laptop provided by the audit office, but this should be possible.
No NDA should be needed.
Depending on the concrete research questions it might be large amounts of data. Also, some possible methods (for example graph autoencoder networks – GAENs) might be challenging to do, so it should be student with good coding and data handling skills using R or Python.
The student will also need to develop some basic domain knowledge of financial auditing and the role of anomaly detection in this line of work.
[1] See for example Schreyer et al. https://www.efk.admin.ch/wp-content/uploads/publikationen/fachtexte/2024_12_a_graph_says_more_than_a_thousand_journal_entries.pdf
[2] www.dfo.no
[3] https://sikt.no/
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. A focus area in the last years is the development and implementationof 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
The object of the current work is to apply machine learning techniques to identify the high temperature softening and melting properties of materials based on images/videos recorded from the sessile drop equipment. The image soft-sensor, developed by this work, will improve data pre-processing and digitalization of the SINTEF wetting lab.
Background:
SINTEF wetting lab focuses on visualizing millimetre-scale samples in sessile drop furnaces. High-resolution pictures will be captured in real-time at high temperatures to evaluate the wetting, softening, melting, reduction, gas production, foaming, and dissolution kinetics, etc. Evaluating thousands or millions of pictures is time – consuming and deviation is verified with various operators. It is challenging to conclude accurate results in an efficient way. In this respect, the use of machine learning is needed to improve the evaluation of the result.
Scope of work:
Skills required:
Data pre-processing skills for image/video using deep neural networks. Super-Resolution technique may be necessary in order to increase the resolution of images.
Reporting:
A written time schedule should be presented at project meetings if necessary. The work should be presented in written form, and common rules of reporting technical work should be met.
Supervisor at NTNU: Prof. Zhirong Yang
Co-supervisor at SINTEF Industry: Dr. Sarina Bao, Dr. Kai Tang
Integrating data from multiple, heterogenous sources represents a challenge. Nevertheless, many use cases in industry demand data integration to unify data access.
The candidate will explore the use of Large Language Models (LLMs) such as ChatGPT, Gemini, Llama, Mistral, or DeekSeek in connection with prompt engineering and fine-tuning. The candidate will compare the proof-of-concept implementation to traditional state-of-the-art methods and assess the potential and risks associated with it.
The project will be co-supervised by Francisco Martin-Recuarda (SINTEF), and Erik Johan Nystad (SINTEF).
For more information see here:https://www.sintef.no/en/digital/master-students/applying-chatgpt-for-data-integration
Rather than HPC enabling AI, what can AI do for HPC?
ML has already been shown to be useful when dealing with lots of parameters when optimizing code, but how do we “co-pilot” paralle programming?
Can "co-iploting also be used for eductainment targening Parallel Programming?
These and other related topics for the motivated student(s) that want to delv into how AI and LLMs can bring the parallel programming world forward are welcome to contact Anne Elster for a chat at elster@ntnu.no
This topic is broad, but we expect the topics will be narrowed after discussions and the student´s interest. The fall project will include a literature survey
Mixing compiler-level precision tuning with fault-injection and software reliability metrics.
Project in collaboration with Leonardo Montecchi (ISSE, NTNU)
Project in collaboration with Politecnico di Milano (IT)
Reliability is defined as the ability of a system to provide continuous correct service. Faults and attacks may affect the reliability of a system, and may have a larger or smaller impact depending on the architecture of the system. Various methods exist for reliability evaluation of systems, for example Fault Trees. We want to investigate how those methods can be applied to machine learning pipelines. This would allow comparing different architectures based on their tolerance to faults.
Project Description
A model is an abstraction of a system that highlights important features of it for a certain purpose, while abstracting the details that are not relevant. Models are used in different ways in software and systems engineering (e.g., UML or SysML diagrams). Some kind of models are used for evaluating quality properties like reliability, safety, or security. As an example, Fault Trees (FT) [1] are a very common kind of model that is used to estimate the reliability of a system.
There has been a lot of work in the literature to design methods to automatically derive fault trees and other dependability [2] models from models of a system or software architecture [3] [4]. The idea is that, from the information documented in the architecture, a lot can be said about how faults can be generated and how errors propagate. A lot of manual modeling effort can therefore be saved. For example, if we know that two components are connected in a client-server pattern, we know that a failure of the server will affect (propagate to) the client.
While those techniques are now established for traditional systems, very few works have addressed reliability models for machine learning pipelines. The objective of this project is to understand what how existing techniques for traditional systems can be adapted to the machine learning context. The idea is to start from some diagram of the architecture of a machine learning pipeline, such as those that can be obtained with TensorBoard [5] or with Netron [6].
This work proposal involves:
The long-term research objective linked to this activity is to define methods that enable system architects to evaluate and compare machine learning architectures with respect to their ability to tolerate faults.
Needed skills
Mandatory
Useful to have (optional)
In such a thesis, we will focus on one of the various AI solutions based on the biometric sensors (eye-tracking, heart rate, EEG) to enable better learing experiences for students. We will also focus on collecting data from students such as eye-tracking, EEG, heart rate, skin temperature, and facial expressions. These data sources provide information about the students from the different points of view and combining the provide better predictions of students' behaviour and their performance.
Requirements: 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.
For details about the different options please contact kshitij.sharma@ntnu.no
The focus of this thesis is to use multiple sensors and artificial intelligence to predict the various performance measurements of pair programmers. Pair programmers usually produce better programming results than the individual programming therefore it is important to understand the factors that contribute to their success. With multiple data sources providing us with information from a diverse points of view, recent works have shown their advantages over individual data sources. In this thesis, the students will use eye-tracking, EEG, heart rate and facial expressions as the data sources.
In the first step, the students will gather data from multiple wearable and pervasive data sources while the participants are pair programming. In the second step the students will develop prediction algorithm using the features from the interaction of the two programmers. Finally, the students will compare the various algorithms and feature sets and write-up the thesis.
Requirements:The ideal candidate will have basic background in machine learning and deep learning algorithms. Programming skills: Python.
Artificial Intelligence for Empowering Women in AI
A previous master thesis has developed a chatbot prototype named EmpowHerAI.
This master's thesis will build on a master thesis that has explored how AI can be intentionally designed to empower women in AI, by deriving a set of actionable design principles and developing a prototype that demonstrates those principles in practice.
In this new master thesis the prototype will be evaluated and re-engineered.
The research question is: How can an artifact be engineered to facilitate the empowerment of women in AI?
In the preliminary phase, the student will run a rapid literature review to examine existing research on AI for empowering women. Then, the study will investigate technology solutions to re-engineer the prototype. Data will be generated through rounds of participatory co-design workshops
This thesis will investigate how AI can be intentionally designed to empower women by embedding participatory design principles. The development of EmpowHerAI.2 will explore how such systems perform across diverse user contexts and investigate how storing and learning from user interactions over time can enable more relevant and personalized adaptation.
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 project is connected to research at NTNU’s COMET research group (Computational Magnetic Metamaterials).
The key focus of COMET is to discover new concepts for ultra-low-power computing systems based on emergent physical phenomena. The research targets new neuromorphic computer architectures outside the standard von Neuman type of operation implemented in new materials (beyond silicon and transistors). Of key interest are magnetic metamaterials called Artificial Spin Ice (ASI), consisting of millions of coupled nanomagnets. ASIs have great potential for massively parallel, ultra-low power computing systems.
The project will mainly be based on simulations using our open-source flatspin simulator for large-scale ASI systems.
If this sounds interesting, please get in touch!
Keywords: unconventional computing, artificial life, neuromorphic computing, artificial spin ice, reservoir computing, bio-inspired systems, complex systems, self-organization, emergence, new technology, neural networks, nanotechnology, material computation, nanomagnetism.
Project description
The goal of this project is to explore new clocking protocols to control the dynamic behavior of coupled nanomagnet systems. This is an interdisciplinary project at the intersection between computing and physics.
To perform computations in ASI, it is necessary to reliably control their dynamics, i.e., how nanomagnets change state over time. In COMET, we have recently discovered a class of field protocols called "astroid clocking". These clock protocols unlock a step-wise and gradual evolution of magnetic states within the ASI, and offers unprecedented control of the dynamical process in both time and space. We have demonstrated astroid clocking of one type of ASI, where the method allows emergent domains in the ASI to be gradually grown or reversed in a step-wise manner. As the domain evolution is driven by an external clock field, the dynamical process can be paused and resumed at will. In other words, the level of control far exceeds what is possible with conventional ferromagnetic materials, and what has previously been achieved in magnetic metamaterials.
However, this is only the tip of the iceberg when it comes to astroid clocking protocols. There is a wide range of topics to explore, e.g., new clock protocols, astroid clocking on other types of ASI systems, methods for protocol design, etc.
This project aims to widen our understanding of astroid clocking. While the primary focus of the project is to explore astroid clocking using flatspin simulations, experiments on real nanomagnet arrays may also be possible.
Links and more information
Material computation is computation that exploits physics directly computation.
Neuromorphic computing are computing systems inspired by properties of the brain and neural systems.
Unconventional computing can be defined as non-von Neumann type of computing, often taking inspiration from nature using design principles like emergence/bottom-up instead of the traditional top-down.
Artificial Spin Ice (ASI) are meta material systems consisting of coupled nano magnets arranged on a 2D lattice, whose collective large-scale emergent behavior has attracted considerable interest as computing substrates.
The project focuses on integrating medical imaging data with natural language processing (NLP) techniques to automatically generate medical reports in an unsupervised manner. The project aims to utilize autoencoders, a type of neural network architecture used for learning efficient data representations. In this context, an autoencoder will be employed to learn a compact representation of both medical imaging data and associated information, such as diagnostic annotations or patient information. Similarly, with the knowledge graph, a structured representation will be created that captures relationships and connections between various medical concepts, which can include image features, diagnoses, treatments, and patient details. The project aims to bridge the gap between medical imaging data and NLP by leveraging the encoded knowledge graph to generate coherent and informative medical reports automatically. This involves transforming the structured information into natural language narratives that accurately describe the medical images, findings, and clinical insights.
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:
Collaboration
In partnership with University of Stirling, Scotland, United Kingdom
Introduction Salmon skin disease constitute one of the main challenges in the modern aquaculture industry, resulting in substantial economic losses each year. Computational biologists are leveraging new and more accurate methods to better understand the problem. The genomic analysis requires accurate annotation of single-cells in salmon samples, which is currently the most significant bottleneck in the uptake of this technology. The manual annotation process is time-consuming, costly, and error-prone. To address these issues, our project aims to leverage state-of-the-art machine-learning software libraries to develop an automated tool for predicting single-cell types in salmon genomic samples.
Project Objectives
Project Description The project will provide the student with a unique opportunity to delve into the world of aquaculture and single-cell data analysis. By harnessing modern machine-learning techniques, the student will contribute to the development of a tool that can significantly expedite the single-cell type annotation process, ultimately aiding in the prevention of salmon skin disease.
Skills and Requirements
Benefits for the Student This project offers a dynamic learning experience in which the student will:
In summary, this project presents a valuable opportunity for a motivated student to make a meaningful contribution to the field of aquaculture while acquiring essential skills in machine learning and interdisciplinary collaboration. We invite dedicated and enthusiastic students to join us in this exciting venture.
This thesis will closely work with Tietoevry Banking and have co-supervision from the company.
This thesis will require two students to work together.
The banking industry is heavily regulated, and compliance with these regulations is crucial for maintaining trust, avoiding legal penalties, and ensuring the smooth operation of financial services. However, the current reliance on manual processes for compliance management is inefficient, error-prone, and costly. This thesis aims to explore and prototype automated solutions to streamline compliance processes, focusing on developing an Agentic AI system and applying it to specific use cases.
Main Task: Develop an Agentic AI System for Compliance Analysis
Objective: To develop an Agentic AI system that automatically analyzes relevant compliance documents for various services or products and assesses adherence to the relevant compliance requirements.
Data Collection and Preparation: We will provide the targeted services or products to be assessed and the relevant compliance documents to analyze.
Agentic AI System:
Application: Apply the Agentic AI system to the application of Visa products and service compliance analysis
This application is to efficiently process compliance documents, extract key information, and generate summaries, insights, actions, recommendations, etc., providing valuable insights for both internal usage and external customers.
Data Collection and Preparation: We will provide the target dataset (e.g., a sample set of compliance documents for some Visa products or services) to be processed and analyzed to develop the prototype.
Data Storage:
Compliance Analysis:
Web UI Development:
Conclusion
This thesis will contribute to the field of compliance management by developing automated solutions that enhance efficiency, accuracy, and cost-effectiveness. The proposed tasks will provide a robust framework for monitoring compliance updates and assessing adherence, ultimately helping banks navigate the complex regulatory landscape more effectively.
The future of Autonomous Driving (AD) is end-to-end approaches (IL, IL+RL), simulators, closing the sim to real gap using Neural Rendering/Reconstruction, TeleDrive over 5G+ if the AD-car need assistance, HD-maps / Digital Twins (DTs) of the infrastructure where you want to drive autonomously, especially in Nordic winter environments (these HD-maps/DTs can also be used in the simulators), crowdsourcing to have an updated Digital Road Twin (e.g. mobile device with accurate positioning and geo-referencing), and automated anonymization of camera data - executed in a way that make the filtered images useful for training and verification of AI models.
Within NAP-lab (= Ntnu Autonomous Perception lab), and associated projects (e.g. MCSINC, MoST etc.), we are working on all these topics (at a state-of-the-art level).
Among other things, our research platform consists of a KIA e-Niro, equipped with NVIDIA DRIVE AGX Orin compute and software stack, 8 cameras, 3 LiDARs, 2 radars, accurate GNSS with CPos corrections, and last but not least, a CAN-bus based drive-by-wire module (DriveKit) that makes it possible to control steering (lateral control) and throttle / brake (longitudinal control) of the car using an API (see the first part of this video). Our old KIA will also be equipped with comma three / openpilot for low-tech end-to-end AD research before the semester starts. On the simulator side we have been using CARLA for quite some time now and will be among the first academic institutions that will have access to NVIDIA DRIVE Sim. In terms of HD-map / Digital Road Twin development we are using NVIDIA Omniverse.
With your help we would like to continue our AD-work within the following areas:
These are just some examples, many other topics exist, and you are welcome to come up with your own project within the AD domain.
For many of the proposed projects, an AI / Computer Vision (CV) background is of course beneficial, but there are also projects where students with other specializations are highly welcome.
We encourage and help to publish papers based on some of the master-thesis work done. For those that are up to it it's also possible to have a research assistant position on the side. It migh also be possible to proceed as a PhD candidate within AD after you finish your master, or as an integrated PhD right before you start on your last master year.
You can work individually, or in pairs of two. Example projects PPTX
Main investigators for NAP-lab related project are Frank Lindseth and Gabriel Kiss
End-to-end Autonomous Driving (but also the more traditional Modular approaches), first in a simulated environment and then test the system in real life using our KIA AD research platform: E2E IL EarlyTest, more info 1, more info 2
Neural Reconstruction (to close the sim-to-real gap): NeRF CARLA EarlyTest, more info
TeleDrive (first and foremost as a backups-system when an AD assistance is needed): TeleDrive KIA EarlyTest
HD-maps / Digital Twins (DTs): Digital Road Twin (from one of our test areas)
Control algorithms for autonomous driving (Sim & Real). More info.
Crowdsourcing (mobile app with accurate geo-referencing for updating the HD-maps++). More info.
Anonymization of camera data: DeepPrivacy2, more info
You can work individually, or in pairs of two.
The project proposes an innovative method for anomaly detection in medical data by combining neural networks with symbolic reasoning techniques. The project explores neural symbolic approach which refers to a hybrid methodology that combines neural networks, which excel at learning patterns from large amounts of data, with symbolic reasoning, which focuses on representing knowledge and making logical inferences. Neural networks can learn intricate patterns from raw data, while symbolic reasoning can help in explaining the detected anomalies and providing contextual understanding
Our planet is changing, now more than ever before. Understanding these changes and how they impact the environment is crucial for preserving the Earth for the coming generations.
Improvements in remote sensing technology and data collection allows us to harvest more data than ever. Hyperspectral remote sensors gather data about electromagnetic radiation reflected off the Earth in nearly the entire spectrum of light emitted from our sun. This data can enable the classification and monitoring of changes in vegetation, agricultural areas, water contents, human habitation, natural disasters and so much more. However, the high dimensionality and redundancy of this data provides a unique challenge for AI.
The project is open for students interested in continuing one of the existing project or applying either a different bio-inspired technique to remote sensing imaging or focussing on another aspect of the imaging challenge.
ongoing/completed projects include
The thesis aims to develop a wildlife monitoring and localizing tool for biodiversity monitoring. Many recent DNN models have been exploited and used on the Serengeti snapshot dataset and Caltech to classify and label wildlife. However, the challenge remains in the successful identification and re-identification of individual species due to numerous factors, including low-quality images, illumination, lighting, background conditions, part of visible animals, or multiple objects within a single image. The thesis objective is to evaluate and develop wildlife monitoring models, including Species Net, for automatic animal detection, localization, and segmentation. The tool will help researchers and biologists in the automatic labeling of animals on a large scale for images captured through camera traps.
Prior knowledge: Machine Learning, Deep LearningSkills: Python programming
Forskjellige Bioinspirert teknikker har vært rettet mot å generere musikk.
En tidlige En annerkjent database og karakteristisk fra Mozart musikk var tilgjengelig for å kunne evaluere at musikk som ble generert hadde liknende karakteristisk.
Det finnes mange måte å generere musikk ved bruk av evolusjonær metoder og studentene er fri til å velge retningen så lenge som det ligger en bioinspirert teknikk i bunnen.
Local learning rules are believed to be more biologically plausible than training models through backpropagation, and may thus overcome model susceptibility to, for example, catastrophic forgetting and adversarial attacks. The article https://arxiv.org/abs/2205.00920 introduces a novel learning rule that learns input patterns without supervision, however, further study is needed before the presented method can be used to solve more challenging machine learning tasks.
This project aims to continue the work towards biologically plausible learning methods with the overall goal to create more robust machine learning models. For instance, the aforementioned learning rule was used for online arbitrary shaped clustering in https://arxiv.org/abs/2302.06335.
Another possibility is to continue the work in the master thesis https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3159736 to make more robust image or sound representations than those currently used in e.g. convolutional neural networks or transformers. Parallelization of the representation methods are also needed, and this is a possibility for students interested in specialized hardware such as FPGA and ASIC to work on.
Please reach out if you find this project interesting: ole.c.eidheim@ntnu.no.
Anskaffelsen av et hyperspektralt bilde (HSI) innebærer å fange flere spektrale bånd innenfor et bestemt bølgelengdeområde. Behandling av HSIer er imidlertid krevende på grunn av den enorme mengden data. Båndseleksjons (BS) metoder er avgjørende for å håndtere utfordringen med høy dimensjonalitet og redundans av HSI data. Selv om det finnes flere unsupervised BS metoder, er det behov for ytterlige forskning knyttet til utnyttelse av romlig informasjon.
Studenter ved IDI har jobbet innen tema, spesielt ved bruk av Particle swarm optimisation or Multiple Objective Optimisation. mange varianter av prosjekter innen tema er tilgjengelig. Det er abre å ta kontakt om du er interreserte i slike biologisk inspirert teknikker og hyperspektralt bilde for å kunne formere din egen vri innen temaen.
relevante tidligere mastersprosjekter:
https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3088496, Publiserte som konferanseartikkel ved IEEE SSCI 2025
https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/2634458
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 are 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. 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
Project Goal: Explore LLMs potential in classifying courses' learning objectives into different levels of Bloom's taxonomy.
Bloom's taxonomy of learning objectives classifies learning objectives into six levels of learning, including remember, understand, apply, analyze, evaluate, and create. Usually, action verbs are sufficient to identify the level of a text (maybe a question item or a learning objective). However, in many cases, action verbs overlap; they appear in varying levels. Therefore, classification based on action verbs only may result in incorrect classification. This thesis idea relates to assessing the potential of LLMs in the identification of text's bloom level.
Objectives/Tasks:
Skills: Python programming
Prior knowledge: Machine learning, Deep learning, knowledge about prompt engineering, and LLMs.
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- Simulating impairments such as blindness, color blindness, dyslexia, or motor impairments- Demonstrating common accessibility barriers in websites and showing how to fix them- 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:
As a small camera moves through an artery, photogrammetry may be used to infer 3D shape inside the arteries. However, there is a tunnelling effect and photogrammetry will typically fail due to a lack of parallax ( displacement or difference in the apparent position of an object viewed along two different lines of sight). . An alternative is to use the image velocity(i.e., optical flow) to computer time-to-collision as long as the camera is moving. Time-to-collision is strongly correlated with depth.
Various biological inspired algorithms or hybrids of such algorithms may be applied within this project.
This project is under a collaboration with University of Waterloo, Canada and provides the possibility to take the pre-project at NTNU and the masters thesis at UofWaterloo. Taking the whole project at NTNU is also a possibility. Currently no data but intenstinal image data is available at Waterloo. Cardio Artery data is to be provided later this summer at Waterloo. If staying in Trondheim an agreement will be reached either with Waterloo or another source for data.
Headaches affect many people. Today, we have access to large data sets describing patients' symptoms, medical measurements, and personal perception. It remains unclear what causes headaches and what remedy is most promising. The master project will explore the use of Artificial Intelligence, particularly Causal Inference, to estimate the causes of headaches in patient records. Causal Inference relies on a scientific model that captures existing knowledge and plausible assumptions. The candidate will get in touch with medical experts to learn more about those. The goal is to construct a causal graph and estimate the effect of random variables on the observed outcome. The outcome can either be modelled as the probability for observing a headache or the probability that a remedy will reduce the perceived pain.
The candidate should have attended courses in about Artificial Intelligence, Probability Theory, and Causal Inference. Ideally, the candidate has a vested interest in the topic of headache. Besides, the candidate ought to have sufficienct experience in probabilistic programming. Experience with medical data is a plus.
These theses topics will be defined in collaboration with CERN for those students interested their Master program. Students participating in the CERN OpenLab summer program are particularly encouraged to apply.
Of oparticular interest these days is GPU computing, AI and/or Big Data.
Customer contact chatbots are increasingly used in the public sector. In Norway, municipalities, NAV, Skatteetaten and other public agencies are employing home-made chatbots as first line of communication with citizens. Some of these chatbots have been criticized because they can be perceived as excluding some citizen groups, and not be developed for the needs of the public sector.
You will do a literature analysis of the use of chatbots in the public sector. Afterwards you will design and conduct empirical study of how citizens use chatbots in their interaction with public agencies. Citizen groups and research questions will be developed in dialog with you.
This research requires a good understanding of platforms, boundaries, self-service, and the needs of the public sector. This understanding will be based on literature analysis and later on empirical data generated by you. As output from this task we expect empirical knowledge, but also recommendations about how to develop and use chatbots in the public sector.
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
AI technologies such as generative AI have the potential to replace humans in some work areas and tasks. Successful deployment of AI requires that humans and AI agents find collaboration models that are satisfactory and provide value. This task will look at ways various types of AI are used in knowledge organizations and how knowledge workers cooperate with AI on a daily basis. Type of AI and type of work practices will be decided together with you.
This is an empirical research tasks. It means that you will use research designs such as case studies or design science to generate new knowledge about the problem area. The outcome of this tasks can be new empirical knowledge, new human-AI collaboration models, or design ideas for new interaction modes.
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.
Optical motion capture systems allow precise recordings of participants’ motor behavior inside small or larger laboratories including information on absolute position. Movement tracking is useful in, for example, rehabilitation settings where tracking a person’s movements during training can give valuable information to provide feedback on whether an exercise is performed correctly or not. Traditional systems often use specialized sensors (Kinect, accelerometers, marker-based motion tracking) and are therefore limited in their area of application and usability. With advances in Machine Learning for Human Pose Estimation (HPE), movement tracking has become a viable alternative for motion tracking.Combining HPE-based movement tracking and EEG can provide the patient with more holistic feedback and help with progress in rehabilitation.
In this study, the students will combine EEG and HPE-based movement tracking, for an existing VR exergame. The EEG can be used to give the patient additional feedback on, for example, her/his attention level, movement intention, or cognitive load. For comprehensive analysis of bio-signals tracked by various sensors, movement and brain data need to be time-synchronized with VR contents. The pipelines that could be used are https://www.neuropype.io/ or https://timeflux.io/ based on LabStreamingLayer. They should work with our EEG, Unity out of the box, however, are not yet synchronized with HPE.
This project is in collaboaration with prof. Marta Molina at the Institutt for teknisk kybernetikk. The study is associated with Vizlab and the needed sensors and the basic training on how to use them (VR headset, EEG equipment) are provided.
When building Large Language Models (LLMs) for Norwegian, acquiring sufficient alignment data is a challenge. This data aims to ensure that the models output aligns with the values and ethical standards of the model creators.
Using synthetic data to countervail data shortages is a normal practice today, e.g., surveyed by Joshi et al. (2024) https://ieeexplore.ieee. org/abstract/document/10423161. However, in a recent report by the RAND corporation about Artificial Intelligence (AI) in the context of war (https://www.rand.org/pubs/research_reports/RRA1722-1.html), it is claimed that “AI cannot make up for a scarcity of appropriate data.”
In this project, the candidate will compare using synthetic and real data for the models created in the TrustLLM and NorLLM projects, both in terms of qualitative and quantiative evaluation. The above problem statement must be operationalized into questions of a form that is possible to answer with scientific experiments of appropriate scope.
Lars Bungum <lars.bungum@ntnu.no> (post.doc.) will be co-supervising the project.
Competence frameworks aim at providing an overview of the competencies that are required in different contexts and/or by different categories of people, for example specifying the (digital) competencies that teachers should have; all the competencies of researchers, …
Though these frameworks are useful at the policy level, they are not easy to be used by individual to reflect on their competencies and to develop an effective professional development plan.
This task aims at developing a tool to help users to playfully self-assess their competencies versus what is recommended by a framework and plan their professional development accordingly. The tool is expected to focus on playfulness, cooperation, and reflection.
The prototype will build around the “ResearchComp: The European Competence Framework for Researchers”, though it is expected to be more general and adaptable to different frameworks.
The task is expected to include design, prototyping and evaluation.
Contact the supervisor to share your ideas and know more about this task
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:
For your own safety, please consider the following expected requirements and you will not hurt yourself while working on this project.
Mandatory:
Desired:
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
Key words: unconventional computing, neuromorphic computing, artificial spin ice, reservoir computing, ,bio-inspired systems, complex systems, emergence, new thecnology, neural networks, nanosystems, material computation.
Material computation is computation that exploit computational properties in underlying physics of material.
Neuromorphic computing is computing systems that are designed and operates in a fashion inspired by the brain and neural system.
Unconventional computing can be defind as non-von Neumann type of computing. Often based on bio/natuur-inspired design principles like emergence/bottom-up instead of the traditional top-down.
Artificial spin ice (ASI) are systems of coupled nanomagnets arranged on a 2D lattice. The collective behavior and large-scale emergent behaviorhas have attracted considerable interest as a promissing metamaterial for compiting.
The unconventional computing and material computing research group at NTNU has shown that ASI is a promissing computational substrates within the reservoiir computing framework. At present time the results are general, basic computational properties. To further demonstrate capability this project will focus on exploring the possibility to adapt the theretical findings to defind real world tasks.
The project will mainly be based on simulations using the flatspin simulator, flatspin is am open-source large-scale artificial spin ice simulator, to simulate ASI reservoirs. Two realworld problems are proposed;
The work may include; finding an efficiant input encoding, defining ASI reservoits with required computatuinal power and defining readout strategies. .
The project is closely connected to the research projects SOCRATES and SpinENGIN where nanomagnets ensembles are explored as a new substrate for computation. We uses nanomagnet ensembles that are Artificial Spin ice. We use Reservoir Computation as computational framework to explore our nanosystems. Within the two projects we produce physical samples that are experimentally tested in synchrotrons and in simulations, mostly simulations use the inhouse simulator flatspin.
Reading material:
flatspin: A Large-Scale Artificial Spin Ice Simulator: A paper presenting our flatspin simulator that most likely will be used (running experiments on the IDUN supercomputer). flatspin simulator and quick start.
Computation in artificial spin ice: A paper illustrating measurements of computational properties in Artificial Spin Ice
Reservoir Computing in Artificial Spin Ice: A paper where Artificial Spin Ice is explored within the reservoir computing framework
Sorry if my txt is a bit hard to get. Anyway, if you are interested in bio-inspired systems, nanosystems, complex systems, recurrent neural networks or just curious on new and existing views on computing please drop me a line and we can arrange a meeting (zoom or hopefully IRL).
This project is a cooperation with a leading industrial partner that is one major player in the field of autonomous driving.
Computer vision for robots, self-driving cars, and also for driver assistance very often assumes very good visibility conditions -- the area is very much dominated by the nice weather conditions that are found when driving e.g. in the Silicon Valley environment.
However, autonomous driving or visual driver assistance in the Nordic Countries needs to deal very often with bad weather conditions: rain or snow, thus wet or snow-covered road surfaces, making vision conditions even worse by the froth or spray caused by vehicles in front of the own one.
This project deals with the exploration what the combination of machine learning and classicial computer vision methods can do in order to detect other vehicles in a distance, road markings, side barriers, and other obstacles as early and as robustly as possible.
One way often being taken in the (sparse) literature about this important topic is to try to improve the quality of the incoming image/video material and subsequently employ 'off-the-shelf' / textbook computer vision algorithms. We do not think that this is the best way to proceed; rather than that, the methods of detecting objects (e.g. in spray, snow, or haze) and the methods for estimating distances and the 3D structure of the environment should consider the afflicted image acquisition process from the very start, and design new algorithms that can cope with the image degradations either by modeling them explicitly (classical approach), or by training learning-based approaches.
The topic of visual environment perception and situation assessment under bad visibility conditions is of course not only of interest in conjunction with cars and road traffic, but equally also for marine applications.
The project can build on an existing database of 'bad weather driving'. If the student(s) working on this project are interested, they can be brought into contact with one of the leading industrial research laboratories for driver assistance and autonomous driving, based in Germany. This may include the opportunity to have a limited guest stay at these labs.
Students who are willing to 'dig deeply' both in the available methods of statistical signal processing as well as in very recent developments and architectures from machine learning / deep learning are invited to contact the prospective advisor in order to explore what this project offers in terms of scientific/engineering challenges and requirements w.r.t. courses and contents that are needed for dealing with this topic.
This project aims at building a mid scale mobile robot using a physical robot that we already have in my group, the LIMO robot by Agilex.ai
https://global.agilex.ai/products/limo
The physical platform (LIMO) is already available at our lab. The project aims to shape this system into an actual robot, interacting with people and/or other robots.
Very probably we aill also procure a 2nd copy of this robot, as it turned out to be versatile in its functionality, and easy to use.
The mobile robot to be developed here focuese on at mid-scale simulation of how mobile robots can interact with each other cooperatively in a dense traffic-like situation, possibly also including interaction with humans. Thus emphasis is not on speed, but on 'intelligence', behaviour, and autonomy.
Note: The final title will include the specific task that the vehicle aims to accomplish
Preliminary work plan: • Specify the goal of the vehicle, e.g., communicating with others to enable collision-free navigation, use of visual and/or positioning information to reach a target as fast as possible, completing domain-specific tasks based on sensor information (for instance, get a package from A to B) • Implement at least one intelligent component, e.g., vision, by using state-of-the-art methods. Possible candidates should be evaluated with respect to their advantages and disadvantages.
Important: the project could also be redefined, in agreement between the supervisor and the student(s) to put most emphasis on the software side of the mobile robot, behaviour, and 'intelligence'.
The 'mobile robot platform' offers a very wide range of possibilities to apply ones creativity for the design and the functionality of the targeted platform. The outcome is not necessarily very 'car-like'; it could also go into the direction of a 'social robot'. The well known R2D2 robot represents nicely the other end of the spectrum of how such a robot could look like.
In case of this project being performed as a Masters project, it would be expected that also at least implementing a part of the robot's intelligent capabilities is included in the agreed project goals. This could be vision, but it could also be on the side of interaction with humans, e.g. some speech or speech understanding capabilities.
The goals to be fixed in the project contract are to a wide degree subject to negotiation with the supervisor, but there will be in any case an engineering challenge in the area of communication, perception, behaviour, or control.
Telecom networks generate massive volumes of multivariate time series data—aggregated counters from RAN and Core components that evolve in time and space. Monitoring these metrics is critical for detecting anomalies, forecasting demand, and ensuring robust network performance. However, existing AI-based models for network data are typically siloed, task-specific, and limited in their ability to generalize across deployment scenarios or leverage rich contextual signals (e.g. support ticket logs, operator feedback, or topological metadata).
The inability to integrate heterogeneous data and context limits the accuracy and deployment of the models. Moreover, current models struggle to scale or transfer across network domains (RAN, Core, Transport network) or Business Units requiring costly retraining or manual tuning.
This thesis explores the development of a foundation model specifically tailored for telecom multivariate time series (MTS) data, designed to support multiple downstream tasks (e.g. anomaly detection, forecasting, imputation) with minimal fine-tuning.
Key contributions include:
Optional directions:
The project will include:
[1] Daesoo Lee et al.. (2023). Vector Quantized Time Series Generation with a Bidirectional Prior Model. In AISTATS. 7665--7693.[2] Yong Liu et al. (2025). "Sundial: A Family of Highly Capable Time Series Foundation Models." arXiv preprint arXiv:2502.00816.[3] Daesoo Lee et al. (2024). “Explainable time series anomaly detection using masked latent generative modeling”. In Pattern Recognition, Volume 156, 110826,ISSN 0031-3203.[4] Liang Y., et al. (2024). Foundation Models for Time Series analysis: A tutorial and survey. Proceedings of the 30th ACM SIGKDD conference on knowledge discovery and data mining.[5] Yue Z., et al. (2022). TS2Vec: Towards universal representation of Time Series. Proceedings of the AAAI Conference on Artificial Intelligence, 36(8), 8980-8987.[6] Behrouz A., et al. (2024). Titans: Learning to memorize at test time. arXiv:2501.00663.[7] Zerveas G., et al. (2021). A transformer-based framework for multivariate time series representation learning. Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 2114–2124.[8] Gu A., et al. (2023). Mamba: Linear-time sequence modeling with selective state spaces. arXiv:2312.00752.
To ease the migration of software applications to the cloud, more and more organisations are adopting a Platform-as-a-Service (PaaS) delivery model. A PaaS environment enables software developers to focus on application design rather than cloud infrastructure aspects, aiming to foster more frequent software deliveries. PaaS solutions can either be provided by an external provider (e.g., Red Hat OpenShift or AWS Lambda), or developed and managed in-house. For example, organisations like NAV (nais.io), Equinor (radix.equinor.com), Finn.no (opensource.finntech.no) maintains and offers their own platform as open source to the public.
This project will investigate the adoption and use of these open-source PaaS solutions in IT organisations and how effective they are in supporting developers to move applications into a cloud infrastructure. The project could also investigate the rationale behind using PaaS as a cloud delivery model and to what extent these open-source projects can be beneficial for other organisations [1,2].
[1] Wulf, Frederik; Westner, Markus; and Strahringer, Susanne (2022) "We have a platform, but nobody builds on it – what influences Platform-as-a-Service post-adoption?" International Journal of Information Systems and Project Management: Vol. 10: No. 1, Article 4.
[2] Wulf, F., Westner, M., & Strahringer, S. (2021). Cloud Computing Adoption: A Literature Review on What Is New and What We Still Need to Address. Communications of the Association for Information Systems, 48, https://web.archive.org/web/20210615090458/https://aisel.aisnet.org/cgi/viewcontent.cgi?article=4300&context=cais
Robots have become important parts of industrial applications as they carry out task in environments unsafe for humans. Still, the interaction remains cumbersome, in particular for multi-robot missions.
Generative AI and AI planning represents promising ideas to create a better experience. Large Language Models (LLMs) can generate instructions for robots, but verifying the correctness of execution plans remains challenging. The candidate will develop an LLM-based model that produces formal mission specifications for robots given input in natural language. The output will be formatted in Planning Domain Definition Language (PDDL).
The project will be supported and co-supervised by Aksel Andreas Transeth (SINTEF). The candidate can get access to robots at SINTEF’s lab to verify the efficacy of the developed models.
You bring problem or method. If it is sufficiently cool and difficult we'll do it.
Counterfactual generation for time series is challenging due to temporal dependencies and causality constraints. While counterfactual explanations are well-explored for tabular and image data, time series presents additional challenges such as ensuring sequential coherence, handling non-stationarity, and generating actionable explanations. Existing counterfactual generation methods often fail to maintain time consistency or produce realistic counterfactuals that are plausible for decision-makers.
This project will develop counterfactual generation methods for time series considering temporal coherence and plausibility, using the HARTH dataset from the HARTH-ML repository. HARTH contains wearable sensor data for human activity recognition (HAR), capturing motion patterns from two three-axial Axivity AX3 accelerometers. While predicting human activities, a counterfactual model could suggest alternative movement patterns to enhance recognition accuracy and interpretability. By integrating Transformer-based and LSTM models from HARTH-ML, this project aims to generate realistic, actionable counterfactuals.
Define and generate valid counterfactuals for time series data without violating temporal dependencies while balancing actionability and plausibility.Develop an approach that outperforms state-of-the-art fidelity, diversity, and comprehensibility methods.
This thesis will focus on developing and evaluating a counterfactual generation method for human activity recognition. The expected contributions include:
Relevant (initial literature):
The project will be supervised by Betül Bayrak and Kerstin Bach. If you have questions, please get in touch with both of us.
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.
Sentence simplification is a process by which complex sentences are rephrased into simpler sentences while retaining the original meaning (See Feng et al.: https://arxiv.org/pdf/2302.11957.pdf).
Traditionally, Machine Translation has been done on sentence-aligned parallel corpora, making the unrealistic assumption of a one-to-one sentence correspondence in natural language. Creating MT systems that preserve linguistic variation is a hard problem, not least with regard to preference for sentence lengths. One way of addressing this problem is to look at sentence splitting with LLMs, i.e., to split longer sentences into more units that preserve meaning as a preliminary step in translation.
This project will begin by acquiring the available monolingual sentence simplification datasets and fine-tuning LLMs before moving to the cross- lingual stage. There are some parallel corpora that can be used to extract one-to-many and many-to-one translation items for training and evaluation of the LLMs.
The candidate is expected to use the NorLLM and TrustLLM models for evaluation, as well as benchmarking against the state-of-the-art models. This constraint limits the scope of source and target languages, which can be relaxed if the candidate has a specific language pair in mind.
Are you passionate about machine learning and eager to contribute to foundational research in this dynamic field? We are excited to offer a flexible and ongoing opportunity for Master’s students to work on thesis projects in basic machine learning research.
This opportunity is open to students who:
We will discuss more details and help you choose among the potential topics. Please email me (zhirong.yang@ntnu.no) for an appointment. Preferably, also send the following to me in your email.
Selected students will work closely with our team to identify and refine a research topic that matches their aspirations and our ongoing work.
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
AI-technology has revolutionized many business processes. In banking, organizations increasingly rely on complicated models such as deep neural nets. These models suggest which products or services to advertise to customers.
While the models might give optimized suggestions, employees need to understand their reasoning. They ask “why should we recommend this product to the customer?”
Explainable AI constitutes a growing research area. The candidate will explore different methods to examine models' output and add explanations.
Sparebank 1 SMN will provide a large data set with up to 200,000 records and 2,000 variables. The candidate will have access to a laptop with which they can access the data and run experiments on a virtual machine.
Stian Arntsen (Sparebank 1 SMN) and Inga Strümke (NTNU) will support the master project.
Neuroevolusjon er en metode som utvikler kunstige neurale nettverk via evolusjonære algoritmer og er inspirert av den naturlige evolusjon av biologiske hjerner. HyperNEAT er en slik metode. Den utvikler mønstre til å bestemme neurale nettverks vekter basert på deres geometri i et substrat. Evolvable-Substrate HyperNEAT (ES-HyperNEAT) har utvidet metoden til å også utvikle nettverkenes geometri. Multi-Spatial Substrate (MSS) utvider HyperNEAT i en annen retning, ved å utvikle forskjellige mønstre til å bestemme vektene i et nettverk som er konstruert over flere substrater.
Ved IDI har studenter utviklet en rammeverk Deep Evolvable-Substrate HyperNEAT (DES-HyperNEAT), som kombinerer de karakteristiske egenskapene til Deep HyperNEAT, ES-HyperNEAT og MSS. Den viktigste innovasjonen med DES-HyperNEAT er utvidelsen av ES-HyperNEAT, fra nettverkskonstruksjon i et enkelt substrat til nettverkskonstruksjon over flere substrater. Den nye metoden utvikler dype nevrale nettverk ved å utvikle og kombinere en rekke substrater.
Noe forskjellige utvidelse av teknikken har vært utviklet av andre studenter og studentene er fri til å utvide teknikken i sin retning for å forbedre teknikken ytterlig.
GlucoSet is a medical technology company developing innovative sensor technology for healthcare applications. Our focus is on creating non-invasive or minimally invasive monitoring solutions that improve patient care while reducing clinical complexity. This project leverages our expertise in sensing technology to address a critical need in cardiovascular monitoring for critical care and surgery patients.
This is not just another AI project - it's about preventing deaths that occur in minutes, not hours.
Cardiac output—the amount of blood pumped by the heart each minute—is a vital measurement that is not used in all critical care or surgery patients who could benefit from it. Today, it's only monitored accurately using very invasive catheters, limiting its availability to the most critical cases despite its clinical value.
Existing non-invasive methods that analyze arterial blood pressure waveforms lose accuracy during hemodynamic instability - precisely when monitoring is most crucial. As demonstrated in the work by Moon et al. (2019), deep learning approaches show significant promise, outperforming commercial systems of such rules-based algorithms. Paired with GlucoSet’s minimally invasive calibration method, we want torevolutionise monitoring in critical care, giving all patients that need it access to acurate and real-time cardiac output monitoring.
This project aims to develop and validate a hybrid Convolutional-Recurrent Neural Network system that:
Although a hybrid CRNN approach shows promise based on previous research, we are open to students that want to explore other deep learning architectures (such as transformers, attention-based networks, or novel hybrid models).
The expected outcome includes a validated algorithm capable of real-time cardiac output monitoring with higher accuracy than commercial systems, especially during critical periods of hemodynamic instability. The work has both scientific value (advancing deep learning techniques for physiological signal processing) and practical impact (potentially bringing cardiac output monitoring to every critical care bed).
The project will utilize public databases (e.g. VitalDB and/or MIMIC-IV), which contain extensive collections of arterial pressure waveforms paired with stroke volume measurements derived from pulmonary artery catheters. These databases include data from diverse patient populations and clinical scenarios.
VitalDB is a comprehensive dataset of 6,388 surgical patients composed of intraoperative biosignals and clinical information. The biosignal data included in the dataset is high quality data such as 500 Hz waveform signals and numeric values at intervals of 1-7 s. MIMIC-IV-Waveform is a subset of the larger MIMIC-IV database, which contains data for over 65,000 patients admitted to an ICU and over 200,000 patients admitted to the emergency department. Patient data is de-identified/anonymized in accordance with healthcare regulations.
The public datasets will be available when the student has passed a simple course in biomedical research and privacy.
Should the patient make inventions that are patentable, delayed publication may be desired so that patenting can be pursued. If student inventions are patentable, the student will be compensated.
This project builds on promising research demonstrating that deep learning methods can outperform conventional algorithms for stroke volume estimation from arterial pressure waveforms. The innovative aspect is developing a system that not only accurately estimates cardiac output but also detects when its own predictions may be unreliable, prompting recalibration.
The concept has been proven feasible in retrospective research (Moon et al., 2019), where a deep learning approach demonstrated superior performance to commercial systems, particularly during hemodynamic instability.
The student will have the opportunity to work at the intersection of artificial intelligence and critical care medicine, potentially developing a solution that could save countless lives.
If you are interested in aspects of modeling software and systems we can agree on a topic of your interest, as long as it is adequate for a Master's project. Topics may range from modeling of architectural aspects of systems, application of formal methods (for example model checking), definition of Domain-Specific Languages (DSL) and metamodels, etc. Just drop me a mail at leonardo.montecchi@ntnu.no to start the discussion.
If you are interested in software and system reliability/safety/security we can agree on a topic of your interest, as long as it is adequate for a Master's project. Topics may range from modeling of reliability/safety/security properties at architecture level, application of formal methods (for example model checking), fault injection, testing, etc. Just drop me a mail at leonardo.montecchi@ntnu.no to start the discussion.
Generative AI (GenAI) is introducing many new concepts in software development. The objective of this project is to define a metamodel and DSL for concepts related to LLMs and generative AI. The DSL will be used to document which GenAI components are used in a software system and how.
A metamodel has been defined in different ways: a model of a model; a definition of a language; a description of abstract syntax; a description of a domain [1]. Metamodels are one of the means of mapping concepts and relations within a certain domain, and they are often used as part of the process of creating a Domain-Specific Language (DSL) [2].
The objective of this project is to define a metamodel and DSL for concepts related to the use of LLMs, and GenAI in general, in software engineering. GenAI is regarded as a disruptive change in software developmen [3]t, which is completely changing the way software is developed and it is introducing many new concepts in software development. However, there is not yet a standardized way to document the use of GenAI in the software architecture of systems.
The long-term research objective linked to this project is to define a methodology to specify and document GenAI components in software and system architectures.
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.
A design system is a collection of documented user interface (UI) elements, visual guidelines, and design principles that other people can use or refer to when designing digital products. Notable examples of design systems are Google’s Material Design and Apple’s Human Interface Guidelines. The main benefits of a design system are (1) improving design consistency across different digital products since it serves as a single source of reference that other people can refer to and (2) reduce the development work since UI designers do not need to design everything from scratch.
As land-based bipedal and quadrupedal robots become more capable, there is a growing need for graphical user interfaces (GUIs) that can be used to operate them remotely. This project will focus on creating a design system that can be used for developing GUIs for operating bipedal and quadrupedal robots.
Key activities in the project include:
This project is part of the OpenRemote project, which aims to create open-source design systems for remotely operated machines. 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), who is also the initial contact person.
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.
Reducing energy consumption in large distributed systems, such as maritime operations, requires effective collaboration among multiple stakeholders. While each stakeholder has their own goals, they also have a common goal, which is reducing the ecological impact of their operations.
This project will focus on designing user interfaces that support coordination between on-board and shore-based operators in terms of environmental feedback. By facilitating more effective information exchange and decision-making, the proposed user interfaces should help both groups of stakeholders to make actions that lead to reduced energy consumption.
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.
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.
As there are many types of cranes, this project will focus on offshore cranes. The key activities in the project include:
Read also: Writing a Master's Thesis in Language Technology
It can be vital both for security and for mental health reasons to identify users at risk in social media. That can be users being at risk of being radicalised (as extremists, school shooters and other types of terrorists, etc.) or of being targeted by predators or extremist organisations, as well as users showing signs of mental health issues (depression, suicide, eating disorders, etc.). This thesis work would experiment with applying various machine learners to social media text. To fully utilise deep learning, a substantial amount of data will need to be gathered, either from previous research or specifically for the task.
A large variety of algorithms have been published to date which allow for 3D surfaces to be compared to one another. These are primarily used to detect point to point correspondences between two potentially similar surfaces. Detecting similarity is the foundation for many things you'd want to do with 3D data, such as an autonomous robot being able to recognise places it has been before.
One problem with all of these methods is that they usually require a scale or size parameter to be defined. The scale of recognising a chair would for example be different than that of a building. Existing work sidesteps this problem by manually setting this parameter. However, it stands to reason that the optimal value of this parameter could be estimated automatically in some way.
The objective of this project is to develop an algorithm that allows the scale parameter of these recognition methods to be selected based on the surfaces being matched. You can think of this like being the autofocus on a photo camera; the resolution of the image is always the same, but the extent to which you are zoomed in or out on a subject depends on what you're trying to photograph. The largest problem to solve is that for two corresponding points on two similar surfaces the estimated scale needs to be the same, even when one of the two surfaces is altered in some way.
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?)
Platform model used by many big tech companies often leads to centralization of power and unfair treatment of users. In this task we want to investigate the concept of platform fairness, the relationship between platform core and periphery, and how new knowledge and design ideas can lead to more fair platforms.
This is an empirical research tasks. Empirical data can be collected from existing platforms such as Foodora or Uber, but also from organizations such as temp agencies. The exact field for empirical research will be decided together with you.
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.
Instersted in Digital Twins (DTs), for self-management of health and decision support or road infrastructure?
Have a look here
Stadig mer av oljeutvinningen på norsk sokkel foregår på svært utilgjengelige felt på store havdyp. Samtidig benyttes stadig oftere ubemannete subsea anlegg som ligger på havbunnen men fjernestyres fra land eller nærliggende rigg. Sensorbasert informasjon (bl.a. temperatur, trykk, seismikk, magnetiske og strålingsegenskaper til bergformasjonen) spiller stadig større rolle i alle fasene: leting ("exploration"), boring ("drilling"), logging og produksjon.
Det (nesten) eneste man har er sensorbasert informasjon - men sensorene er upålitelige. De slites (levetid typisk et drøyt år), de er ikke kalibrert osv. Så hvordan skal operatørene ta gode beslutninger gitt høy usikkerhet? Hvilken rolle har nye IT verktøy?
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
The Approximate Nearest Neighbor (ANN) algorithm is a popular algorithm used in machine learning and data mining. It is used to find the nearest neighbor of a given point in a high-dimensional space. However, as the size of the dataset grows, the performance of the ANN algorithm degrades. This is because the ANN algorithm requires a lot of memory and computation power to work efficiently.To address this issue, researchers have proposed distributed versions of the ANN algorithm that can scale beyond the 1 billion vector mark. However, these distributed versions are not efficient enough to handle large datasets. Therefore, there is a need to build a distributed version of an ANN algorithm that can offer efficient scaling beyond the 1 billion vector mark.Your thesis will explore how to build such a distributed version of an ANN algorithm that can offer efficient scaling beyond the 1 billion vector mark. You will investigate different approaches to building such an algorithm and evaluate their performance on large datasets. You will also explore how to optimize the performance of the algorithm and how to make it more efficient by using different strategies for cross machine partitioning and communication.
This project will require extensive programming, and students should have excellent skills in C++ and Python to take this on. Furthermore, experience and good capabilities in performance evaluation and understanding of detailed performance characteristics of computers will be required.
The aim of this thesis is to explore the current practices in Norway regarding diversity aspects (including but not limited to: identity, gender, race, ethnicity, neurodiversity, sexual orientation, age, and physical abilities) within software development processes and products. The research will be conducted through empirical software engineering methods.
The student(s) will:
Examine how various companies incorporate diversity into their organizations and projects.
Investigate how software development teams reveal diversity in their projects and products.
To achieve this, the student(s) will:
Conduct interviews with customers and students (autumn semester).
Analyze materials produced by groups in the TDT4290 course.
Perform report and software analysis.
Organize workshops with companies and students to validate their findings.
The goal is to understand how diversity aspects impact software development teams and identify effective guidelines to disclose diversity in software development processes.
Cico, O., Jaccheri, L., Nguyen-Duc, A., & Zhang, H. (2021). Exploring the intersection between software industry and Software Engineering education-A systematic mapping of Software Engineering Trends. Journal of Systems and Software, 172, 110736.
A.R. Gilal, J. Jaafar, S. Basri, M. Omar and M. Z. Tunio, "Making programmer suitable for team-leader: Software team composition based on personality types," 2015 International Symposium on Mathematical Sciences and Computing Research (iSMSC), Ipoh, Malaysia, 2015, pp. 78-82, doi: 10.1109/ISMSC.2015.7594031.
L. Gren and P. Ralph, “What makes effective leadership in agile software development teams?” in Proceedings of the 44th International Conference on Software Engineering, 2022, pp. 2402–2414.
Y. Wang and M. Zhang, “Reducing implicit gender biases in software development: does intergroup contact theory work?” in Proceedings of the 28th ACM Joint meeting on european software engineering conference and symposium on the foundations of software engineering, 2020, pp. 580–592
Gunatilake, H., Grundy, J., Hoda, R., & Mueller, I. (2024). The impact of human aspects on the interactions between software developers and end-users in software engineering: A systematic literature review. Information and Software Technology, 107489.
https://sbs.idi.ntnu.no/
Problem statement:
To support the end-to-end (E2E) learning of a driving system, a simulation environment should ideally:
As of today, there is no open-source simulator that truly covers all three properties:NuScenes, Waymo Open, KITTI and other pre-recorded datasets are limited to offline/open-loop evaluation.The CARLA simulator simulates a proper closed-loop environment, but its scenarios are scripted and not very complex.Finally, the nuPlan simulator supports both closed-loop evaluation and complex real-world situations, but sensor inputs are only available in the open-loop setting, not in closed-loop simulation.
Approach:
Recently, novel view synthesis methods (Neural Radiance Fields, 3D Gaussian Splatting) have become very successful.Such methods can generate high-quality RGB images of a scene from previously unseen positions and view angles. This is precisely what nuPlan is missing.
The goal of this project is to explore, how such view synthesis methods can be used to extend the nuPlan simulator by giving it the ability to synthesize camera-inputs on the fly.
Sound interesting? Don't hesitate to reach our.
GlucoSet is developing innovative continuous glucose monitoring systems for ICU patients. Our core mission is to reduce mortality and infection risk through better insulin management in critical care settings. This master's thesis project will build upon our glucose monitoring platform by addressing the critical and often overlooked challenge of potassium level monitoring - a life-threatening complication of insulin therapy that currently lacks continuous monitoring solutions.
Potassium imbalance (dyskalemia) is one of the most immediately life-threatening conditions in ICU settings. While insulin therapy is critical for glucose management, it dramatically affects potassium levels:
The critical challenge is that current methods cannot provide continuous monitoring of this potentially fatal condition. When potassium levels become dangerously high, clinicians may have mere seconds to respond - making this one of the most time-sensitive monitoring challenges in critical care.
This project aims to develop a deep learning solution that can non-invasively and continuously monitor potassium levels through ECG analysis. Building on groundbreaking research by Chiu et al. (2022), we seek to advance this technology by solving the "cold start" problem - enabling accurate monitoring for new patients without requiring prior examples of their abnormal potassium states.
The scientific challenge lies in creating a robust model that can generalize across patients while maintaining high accuracy. The practical challenge is developing a system that integrates seamlessly with existing ICU workflows and various ECG lead configurations.
Expected outcomes include:
The project has immediate life-saving implications while advancing the scientific understanding of how physiological signals can be interpreted using deep learning approaches.
The project will utilize the MIMIC-IV-ECG dataset, a publicly available critical care database containing approximately 800,000 ECGs from around 160,000 patients, with 12-lead, 10-second recordings. Many of these ECGs are paired with potassium measurements in the MIMIC-IV dataset. The student will need to determine what percentage of ECGs have corresponding potassium measurements within an appropriate clinical timeframe, and develop methods to handle the temporal relationship between measurements.
The MIMIC datasets are de-identified and freely available to researchers after completing a data use agreement. The student will need to complete the necessary training and agreements to access this dataset, which GlucoSet will assist with. As this is a publicly available research dataset with established protocols for academic use, we do not anticipate delays in data access.
The data is pre-processed and de-identified, making it appropriate for academic research, although the student will need to pass a simple online test/course to gain access. Should the patient make inventions that are patentable, delayed publication may be desired so that patenting can be pursued. If student inventions are patentable, the student will be compensated.
This project addresses one of the most urgent unmet needs in critical care. Because potassium imbalance can cause death more rapidly than almost any other condition in the ICU, a successful solution would represent a landmark advancement in patient safety.
With the increasing reliance on mobile applications in daily life, concerns about their energy consumption, data usage, and environmental impact have grown significantly. Mobile apps are ubiquitous, but their development and usage can have significant impact on sustainability (both social and environmental). Mobile apps contribute to carbon footprints through intensive CPU processing, network activity, and inefficient coding practices, often resulting in excessive battery drain and resource waste. Developing eco-friendly mobile apps is crucial to reducing these impacts while enhancing performance and user experience.
This MSc thesis aims to develop practical guidelines for designing mobile applications that minimize energy consumption and resource use throughout their lifecycle. This thesis aims to bring a comprehensive evaluation of the areas/ characteristics of mobile apps that contribute to reduce impacts on sustainability.
Suggested research question for this study could be: How do software development practices contribute to sustainable mobile applications? Which design principles and coding practices/ architectural strategies can developers adopt to minimize the energy consumption and resource use of mobile applications?
The research starts with an extensive literature review to existing research on eco-friendly mobile app development and practices. Then, the thesis will carry out a benchmark analysis of existing apps and analyze energy consumption patterns across a sample of widely used mobile apps (e.g., social media, navigation, and gaming apps) using existing tools. Based on qualitative interviews with developers and industry experts (mobile app developers, UX designers, and software engineers), data will be gathered to provide insights into current development practices and the challenges of balancing performance with energy efficiency. Based on findings from the benchmark analysis and interviews, expected outcome of the thesis is to propose practical guidelines that will empower developers to design mobile applications with improved energy efficiency, reducing their environmental footprint. The will be part of the activity at CESICT - Center for sustainable ICT
Since its maiden release into the public domain in 2022, ChatGPT garnered more than one million subscribers within a week. The generative AI tool ⎼ChatGPT took the world by surprise with its sophisticated capacity to carry out remarkably complex tasks. The extraordinary abilities of ChatGPT to perform complex tasks within the field of education have caused mixed feelings among educators as this advancement in AI seems to revolutionize existing educational praxis. This topic will investigate the opportunities and challenges of generative AI (through implementing a case study) and offer recommendations on how generative AI could be leveraged to maximize teaching and learning. The goal is to identify how these evolving generative AI tools could be used safely and constructively to improve education and support students’ learning.
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
With the rise of functions as a service (FaaS) due to the broader availability of cloud computing workloads, we face an evergrowing body of software written in so-called high-productivity languages, such as Python or Javascript. While we have a somewhat detailed understanding of how compiled software performs on modern systems, these interpreted programming languages pose unprecedented challenges. All these languages require a runtime consisting of compiled code which interprets the actual logic and translates it on the fly into an intermediary representation, similar to the bytecode representation of C# or Java.
The system sees the runtime as regular code and thus uses the hardware structures designed to handle code. However, the actual functionality of a program written in an interpreted language is considered data from the hardware perspective. Therefore, it is handled by hardware designed to manage data. The effects of that behaviour on the microarchitectural state of a processor, the memory hierarchy, and the entire system's performance need to be better understood.
To achieve the level of understanding needed, we require better tooling to measure the behaviour of such workloads throughout the computing pipeline. We need more than the currently existing tools to track the data required to assess those effects. The first part of the project is concerned with designing and exploring the space for such a tool that can precisely pinpoint which cache lines contain actual data that the interpreted program is operating on and which cache lines have code to be interpreted by the interpreter.
While the tool helps us track the information in the front end, we must understand its effects on the decoupled front-end and back-end aspects. While the back end is most likely only indirectly affected by the challenges that interpreted languages pose (e.g. stalling introduced by reducing the effectively available L1D cache), the front end is expected to be affected very directly, starting with the inefficient use of caches and the wrong assumptions of the replacement policies, to the prefetchers that are geared towards data or instructions, to the branch predictors, that is lost because the interpreters branching behaviour mostly depends on data and thus is hard to predict.
Understanding such problems' behaviour is of utmost importance to address those challenges. With Python and NodeJS being the two single-most used frameworks for the Function as a Service paradigm and PHP still being the driver for many sizeable open-source back-end software projects, understanding and optimizing hardware towards these types of programs is vital. Understanding the execution allows us to propose new methods that not only ensure the efficient execution of such programs to meet performance requirements but also targets the energy-efficient execution of workloads on cloud computing infrastructure.
This thesis aims to explore the use of DiskANN for batch processing and searching over immutable data. The first step will be to build a good Python interface for DiskANN that can be used to experiment with the algorithm. The thesis will then investigate how to optimize the performance of DiskANN for batch processing and searching over immutable data. This will involve exploring different strategies for cross-machine partitioning and communication.The thesis will also evaluate the performance of DiskANN on large datasets and compare it with other state-of-the-art approximate nearest neighbor algorithms. The results of this evaluation will be used to identify areas where DiskANN can be improved and to suggest directions for future research.This project will require extensive programming, and students should have solid experience in C++ and Python to take this on. Furthermore, experience and good capabilities in performance evaluation and understanding of detailed performance characteristics of computers will be required.
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.
Requirements:
[MLIR] https://mlir.llvm.org
[architecture] https://arxiv.org/abs/2301.03904
Most organizations are currently undergoing different types of digitalization and digital transformation. The majority of these digitalization processes are initiated and led by top management, in collaboration with external consultants and technology vendors. This means employees and their in-depth knowledge of their work practices are often excluded from these processes because the employees don't have the time, the skills, or the autonomy to participate effectively. The consequence is often failed or costly projects.
In this task we ask you to investigate successful and failed attempts to employee participation in digital transformation, understand the challenges, and recommend best practices for future digitalization initiatives.
This is an empirical research task that is grounded in latest research in digital transformation. We expect that you employ rigorous methods to analyze past research, plan research studies, and conduct empirical studies in organizations that are undergoing digital transformation. The specific case organization and the specific focus of the task will be further developed in collaboration with you.
This task requires that you have a good understanding of, and are interested, in empirical qualitative research. Working language for this task is Norwegian. The thesis can be written in Norwegian or English but we recommend English. Please contact Babak before you select this task.
Supervisors: Michail Giannakos, Giulia Cosentino
Place: LCI Lab: https://lci.idi.ntnu.no/
Suitable for: Two students
You can see a video of the current prototype here: https://www.youtube.com/watch?v=3tpqfkSDkE8
IntroductionThe field of multisensory technology allows the learner to interact beyond usual input devices (e.g., keyboard and mouse) with new forms of natural user interfaces such as full body interaction, gestures, vocal, etc., and provide multisensory stimulations (e.g., through sounds, lights, visuals) to support users' learning. Moreover, the number of sensors provided by the multisensory technologies supports the collection of multimodal data that add sensemaking and predictive powers to previous forms of analysis. The goal of this project is to explore how Motion-Based Technology (MBL) and Gen AI capabilities (e.g., LLMs and Multimodal LLMs) may facilitate children’s meaningful STEM learning.
Thesis DescriptionIn the first step, the student needs to review the literature and familiarize themselves with the adaptive learning models, databases and processes, and MMLA, as well as the context of multisensory technologies. Then, the candidate, supported by the best practices found and adapted from the literature, will define the model able to suggest the activity and interaction the learner will have to perform with a given content and interaction paradigm based on the MMLA. Overall, the project aims to: 1) identify aspects of Motion-Based Technology (MBT) and analytics that support children’s learning, with a focus on math, 2) develop a set of practices, functionalities and technological interfaces that materialize those aspects, and 3) evaluate and refine those practices and the technological interfaces.
RequirementsThe ideal candidate will have a background in data science and modeling, and experience with using/integrating LLM technology to applications. Solid back-end programming skills (Python or C# or JavaScript) and an interest in hands-on development and experimentation is also a requirement.
Programming skills: Python or C# or JavaScript
This project explores the significant challenges and proposes innovative solutions within the field of Retrieval-Augmented Generation (RAG) techniques, particularly focusing on domain-specific applications. One of the primary challenges addressed is the development of robust embedding vectors that accurately capture the nuanced language and terminologies unique to specific domains. This involves enhancing the algorithms that process and understand textual content, thereby improving the model's ability to integrate and utilize domain-specific knowledge effectively.
Another critical aspect of this research is improving the precision of citations generated by these models. As retrieval-augmented generation techniques pull information from various sources to produce content, ensuring the accuracy and relevance of citations is paramount. This project aims to refine the retrieval mechanisms and citation algorithms to increase the reliability and verifiability of the generated texts.
Through a combination of empirical research, algorithmic enhancements, and practical testing, this thesis aims to advance the field of retrieval-augmented generation by addressing these key challenges.
AI-powered assistants are increasingly used in educational and professional contexts to support learning, task automation, and decision-making. In software project management education, providing timely, relevant, and high-quality feedback is essential but labor-intensive. With the advent of Large Language Models (LLMs), techniques such as prompt engineering, retrieval-augmented generation (RAG), and few-shot learning present promising opportunities to improve feedback mechanisms without requiring full retraining.
InnSpill AI, an EdTech startup developing Learnix, aims to integrate advanced LLM-based assistants into software engineering learning platforms. The startup has collected real-world data from university-level software project management courses, providing a rich context for applied research.
The goal of this project is to explore and evaluate how prompt engineering, RAG, and few-shot learning can be used to improve the relevance, clarity, and personalization of automated feedback for software project management tasks (e.g., sprint planning, risk analysis, reflection logs).
Design Science Research Methodology (DSRM) to iteratively design and evaluate the feedback-enhancement system. This thesis will be conducted in collaboration with InnSpill AI, which provides access to data, expert supervision, and platform integration opportunities.
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/
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.
Svermrobotikk blir mer og mer relevant for moderne samfunn ettersom teknologien legger til rette for å løse mange problemer både billigere og mer effektivt enn andre alternativer. Svermroboter tilbyr løsinger på problemer ved å bruke grupper av selvstyrte roboter. Robotene viser intelligent adferd som en egenskap som oppstår som følge av summen av det de gjør. For å sørge for at kontrollmekanismene er tilpasningsdyktige og kan lære seg ny adferd blir kontrollmekanismene kunstig utviklet ved hjelp av metoder inspirert av naturlig evolusjon.
Forskjellige evolusjonær sverm robotikk prosjekter er tilgjengelig. Noen eksempel av tidlgiere prosjekter:
Studentene står fritt til å vri en av disse eksisterende prosjektene i sin ønsket retning eller til å utarbeide egen prosjekt innen temaet.
In AI and ML, several methods rely on stochasticity or randomization: mutation and crossover in evolutionary algorithms; dropout and stochastic gradient descent in deep learning; stochasticity in stochastic local search; and randomization in systematic search. Evolutionary algorithms (EAs), which we study here, are competitive in solving many computationally hard problems while modeling important biological phenomena. Further, EAs are interesting in that they can studied formally, for example by means of Markov chains.
Project Focus
The goal of this project and thesis is to study the theory and application of EAs. The following research questions are examples - specifics depend on finding a topic of broad and common interest. First, there are opportunities to improve EAs, with an eye to specific applications. Can this done by combining different heuristics and adaptive methods, perhaps including concepts from algorithms for stochastic local search?
Second, the computational cost of machine learning (ML) can currently be high, and the use of EAs for ML, or evolutionary ML (EML) is no exception. Maybe Markov chain-based analysis [1] and/or distributed computing can be the basis for EA-based ML methods that better handle massive and complex datasets while reducing computational cost dramatically?
Third, the theory and formal methods communities have studied parametric Markov chains, but interactions with the AI and ML communities have been limited. Perhaps this will change, now that trustworthy AI is coming to the forefront?
[1] Ole Jakob Mengshoel, Eirik Lund Flogard, Tong Yu, and Jon Riege. 2022. Understanding the cost of fitness evaluation for subset selection: Markov chain analysis of stochastic local search. In Proceedings of the Genetic and Evolutionary Computation Conference (GECCO '22), pages 251–259. https://doi.org/10.1145/3512290.3528689
Notes
In this project, the joint interests of (potential) sponsor, advisor and student(s) come together. Typically, the project will be based on the problem described by the sponsor AND previous research by Prof. Ole Jakob Mengshoel AND interests of students. If only one or two of these are present, there is no basis for a project.
+++++
Please send email(s) to (potential) advisor(s) and/or sponsor(s) if you're interested in this project.
Ole Jakob Mengshoel is more or less following Keith Downing's selection process for master students - please read this:
https://folk.idi.ntnu.no/keithd/ai-masters/kd-masters-selection.html
Et mulig prosjekt innen tema, bygger seg på en eksisterende prosjekt gjennomført av en tidligere samarbeidspartner.
Metabolisk syndrom er et enormt helseproblem som påvirker stadig fler i den vestlige verden. Desiste tiårene har gjort oss mer stillesittende og gitt oss enklere tilgang på ultraprosessert mat samtidigsom jobbene våre har blitt mindre aktive. Resultatet er at stadig fler av oss havner under kategorien metabolsk syndrom. Dette øker risikoen for å utvikle hjerte og kar sykdrommer samt diabetes type 2.
Studenter kan justere oppgaven til anvendelse av interest og ta i bruk en annen biologisk- inspirert metode
Typisk prosjekt
If you have a clear idea of an exciting topic you want to work on in HPC (high-performance computing), including Parallel Computing, Quantum Computing simulations and Cloud Technologies, contact me (Anne Elster elster@ntnu.no) to see if we are a fit and how feasible you project is as a fall project and/or Master thesis.
Topics include investgating new parallel algorithms, GPU features, including testing NVIDIA B100s,and other new HPC systems.
I am aslo interested in developing edutainment for teaching parallel programming and parallel computing!
Prorject in collaboration with the Department of Biology of University of Pavia (ITALY)
This project intends to develop and evaluate the combination of genome-to-image or genome-to-text encoding solutions, deep learning architectures, and explainable-AI methods in capturing higher-order interactions within datasets of genetic variants.
Dedicated methods are needed to capture interactions, but they are computationally demanding and severely limited in detecting higher-order (and long range) interactions, which most likely explain the complexity observed in biological organisms. Combined to deep-learning, post-hoc explainability methods have the potential to add discovery capability to DL architectures: by scoring the importance of input features which contribute to the outcome of a deep neural network, xAI provides a way to identify relationships in the data that other methods might not be able to reveal. These approaches, however, can only capture the information which is represented in the input data: while methods have been proposed to encode genome information, they tend to represent short-range context, functional to the applications the methods have been developed for (identifying variable regions, predicting gene expression, predicting protein binding regions, to mention a few). We have worked until now in developing representations of the genome which might allow deep learning architecture to capture a wider context and therefore aid us in the discovery of interacting variants associated to phenotypes.
Beyond the encoding methods, we face additional elements which an AI-oriented student might find challenging and fun to work on, such as:
- Data dimensionality: encoding large genome is compute-intensive and computational solutions are needed to speed up the conversion of genetic data into manageable tensors for deep learning experiments
- Representation of sparse events: some of the genetic variations we need to represent are quite rare, and therefore any deep learning architecture might encounter challenges in identifying meaningful patterns in an otherwise invariant background. Solutions to address this issue, might prove invaluable
- Experimenting with different architectures, appropriate for the two issues described, with learning / accuracy profile corresponding to the expectations for the Human genome (not everything in a disease is genetically determined)
- Appropriate explainability methods, for the goal of discovering new genetic determinants of disease for this project
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.
The project will mainly be based on simulations in mumax3 as well as our open-source flatspin simulator for large-scale ASI systems.
The goal of this project is to tailor the switching characteristics of nanomagnets for computation. This is an interdisciplinary project at the intersection between computing and physics.
ASIs are currently being investigated as a promising substrate for material computation. To this end, the COMET group recently published a novel method to interact with ASI systems known as “astroid clocking”. Astroid clocking is a powerful way to supply input to and control the dynamics of an ASI.
Astroid clocking exploits the shape of the nanomagnet switching astroids’ within the ASI. As such, it is of great interest to know what astroid shapes are possible and how they can be achieved. There is little published knowledge and data on what astroid shapes are possible and how to achieve them, aside from a few canonical examples.
Elongated nanomagnets behave as binary spins, i.e., with two stable magnetization directions. A nice feature for computing, we can treat the state of these magnets as binary, “1” or “0”. A magnet will flip between these two states when a strong enough magnetic field is applied. The required field strength depends on the angle at which the field is applied, and this relationship is described by the magnet’s switching astroid.
The switching astroid of a nanomagnet is governed by the magnet’s shape and size. This aim of this project is to investigate what astroid shapes are obtainable. This could be exploring the possibilities with an evolutionary novelty search sort, or otherwise. Alternatively, a method of searching for a suitable magnet shape given a target astroid shape could be developed. In either case, the project involves running micromagnetic simulations of magnet shapes to determine their respective astroids.
Kunstig immunesystemer har mange varianter men en kjerne mekanisme er ‘recognition regions’ (RRs), hvor en ‘antibody’ kan klassifisere ‘antigens’ i sitt område. Tradisjonalt er slik RRs rund men for mange applikasjoner med høydimensionale rom blir deknignsgrad for lite med slike RRs. En ny løsning var foreslått av IDi studenter som heter ‘open’dimensjoner'. Videre undersøkelse av denne nye mekanismer trengs mht deknignsgrad og muligheter mot flere applikasjonsområder.
Studenter står fritt til å velge en fortsettelse av denne oppgaven eller en relatert pågående oppgave eller ta sin egen vri på oppgave. en slik vri kan og være noe annet videre utvikling av et eller annet kunstig immunesystem.
relaterte student oppgaver:
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 focus of this thesis is to develop a system to help the novices in programming while debugging. One of the ways to provide help is to learn from the expert about how to look at the program while finding the bugs in the code. This way of providing help is called Expert’s Movement Mapping Examples. Most of the efforts in this direction include the use of expert’s gaze in the problem space. In this thesis the student(s) will exploit the use of dialogue as well as the gaze of the expert.
Thesis DescriptionIn a first step, the student(s) will design and implement the gaze-enabled feedback tool. Afterwards, they will conduct a small user study in order to test the usability of the system with a small 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. Solid programming skills and an interest in hands-on development and experimentation is also a requirement.Programming skills: Python/Java.
Online meeting platforms are beginning to offer their raw image and sound data for processing via SDKs; for example Zoom:
https://developers.zoom.us/docs/video-sdk/linux/raw-data/
The idea behind this project is to develop facial processing tools based on Zoom's SDK.
Knowledge: Linux, Python, C/C++
Courses: TDT4195 (Visual Computing Fundamentals), TDT4230 (Graphics & Visualization), or equivalent.
Supervisor:
Prof. Theoharis Theoharis, Dr Antonios Danelakis IDI, NTNU theotheo@ntnu.no
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.
In this project, the aim is to address the challenges of analyzing medical images while maintaining patient privacy. In this approach, instead of centralizing all medical data in a single location, the project employs federated learning, a decentralized machine learning technique. In a federated learning setup, multiple medical institutions or entities collaborate to train a shared machine learning model without sharing raw data. Each institution retains control over its data, and only model updates are exchanged and aggregated among the participants. This approach ensures that sensitive patient information remains localized, minimizing the risk of data breaches and privacy violations. The project's focus on medical image analysis indicates that it's aimed at tasks like diagnosing diseases, detecting anomalies, or segmenting regions of interest within medical images. By using a collaborative federated learning approach, the project combines the benefits of data sharing for model improvement while upholding strict privacy standards required in the medical field.
Large language models form the basis of almost all currently topical AI research, making it vital to identify and rectify different types of demographic biases in those models (be that bias based on gender or sexual identity, or on cultural, ethical or social background, etc.). 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. There is still a lot of room for improvements though, both in identifying and quantifying bias, in developing dibiasing methods, and in defining bias as such.
Projects will be based on the students chosen GA towards portfolio optimisation or some other element of finance, chosen by the student.
One potential avenue is continuation of an earlier masters project based on multi-objective differential evolution for VaR portfolio Optimisation.
Wildfires pose a significant threat to ecosystems, communities, and infrastructure. This project focuses on the development and enhancement of wildfire detection and early warning systems using AI and computer vision. By leveraging a combination of remote sensing data, sensor networks, machine learning, and artificial intelligence, the study aims to create a robust and proactive solution for early wildfire detection. The research delves into the challenges of accurately identifying wildfire signatures from various data sources, optimizing real-time data processing, and designing efficient alert mechanisms to enable timely response and mitigation efforts. The outcomes of this research contribute to minimizing the impact of wildfires on lives and the environment through advanced fire detection strategies.
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
For Eirik og Jørgen: Syntetiske data med SMN1
Existing Generative AI facilitates transcribing interview. Currently, journalists take those transcriptions and write articles based on their information. The process involves recurring steps. Journalists must summarize the main messages, and possibly provide quotes that reflect what the interviewee said. It is conceivable that an AI can take over some of this process.
The candidate will obtain a data set with interview transcriptions and published articles from Adressavisen. Using these resources, the candidate will fine-tune a set of Large Language Models (LLM) to draft articles based on interviews. Possible models include GPT, Llama, Mistral, Gemini, or DeepSeek. The interviews and articles need to be in Norwegian presenting an additional challenge. Journalists at Adressavisen will evaluate the quality of the most promising models in a user study.
The project aims to advance the understanding and interpretation of three-dimensional (3D) point cloud data through segmentation techniques. Point cloud data represents objects or scenes in 3D space using individual points, often collected from technologies like LiDAR sensors or depth cameras. Semantic segmentation in this context refers to the task of labeling each point in a point cloud with a specific class, such as "car," "tree," or "building." This enables the identification and categorization of objects within the 3D environment. Panoptic segmentation, on the other hand, extends this concept by not only labeling object instances but also recognizing 'things' (e.g., individual objects) and 'stuff' (e.g., surfaces or background). The project aims to develop a comprehensive approach that bridges the gap between semantic and panoptic segmentation for point cloud data. This involves creating models and algorithms capable of not only recognizing object categories within the point cloud but also differentiating between individual instances of those objects and understanding the overall context in which they exist. The potential applications of this project are significant. It could contribute to advancements in autonomous driving by enhancing the ability of self-driving vehicles to perceive and react to their surroundings accurately. Additionally, it could be useful in various fields where 3D data analysis is critical, such as robotics, augmented reality, urban planning, and environmental monitoring.
The immune system is arguably one of nature's most highly adaptive, distributed and self-organising systems. It has the property of being able to recognise anomalies --- something that deviates from the common rule. In 2018-2019 two masters students proposed a novel hybrid classification algorithm MAIM, combining such features of the Immune Systems with an Island Model Genetic Algorithm (IGA). The preliminary results achieved are promising and resultet in an international publication. The goal of this project is to build on this foundational work and investigate the many avenues available to further extend/refine this novel algorithm.
Contact Pauline haddow, pauline@ntnu.no for further information
As a result of increasing digitalization, IT tools are used in multiple contexts by users with multiple backgrounds, and are intertwined in complex ways with everyday practices. The complexity of the context of use arises a number of ethical issues for software developers and users, including e.g., algorithmic biases, privacy of personal data, addictive design, concerns about sustainability,…
This specialization/master project aims at the development and evaluation of a game to facilitate reflection and discussion on ethical issues related to IT tools among informatics students.
Previous work has been done in the group about teaching about ethics to informatics students and provides a good starting point, still giving freedom to shape your work.
Contact the supervisor to share your ideas and know more about this task.
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 game to help workers to understand the space of possibility of new technologies in their workplace and their impact. Focus will be on systemic and critical thinking. Students are welcome to define, in cooperation with the supervisor, specific areas of interest, with respect to the learning objectives of the game, and game genre and technology.
AI tools are increasingly used in different workplaces. This project focuses on the use of AI for supporting creativity, with focus on ethical and responsible use. More specifically, the task is centered around the design and evaluation of a game to learn about how to use AI for supporting creativity and innovation.
Students are welcome to define, in cooperation with the supervisor, specific areas of interest, with respect to specific target groups (e.g. specific workplaces), the learning objectives of the game, and game genre and technology.
The specialization project is expected to focus on understanding how AI can be used for supporting creativity. The work can then be continued with a master focusing on the design and evaluation of the game.
SINTEF Digital, Cognite, Kongsberg, NorwAI: In hybrid AI the aim is to combine machine learning and physical knowledge. This is particularly relevant for industrial applications of AI and machine learning on physical systems. Through the SFI NorwAI we work on hybrid AI together with partners in the energy sector. We have two use cases focused on predictive maintenance of wind turbines and on virtual flow meters in oil and gas.
Problem Description:
Simulation based inference is a likelihoo-free Bayesian inference method where one uses machine learning to approximate the likelihood function based on simulation data. This allows for Bayesian inference when no analytic model exists. The method naturally performs best in cases of stochastic models. The project can either have an applied or a theoretical focus. For the applied focus, the goal is to combine existing commercial black box simulation software and real data from wind turbines and compressors to speed up model identification (which input parameters are required for the simulation to match the data) and quantify the model uncertainty. For the more theoretical focus, the goal is to study different machine learning models for probability distributions and how they can be constrained to meet certain physical constrains.
The overall learning outcomes will be
The project can be adapted to students from several departments such as Computer Science, Mathematical Sciences or Engineering Cybernetics.
Data Description: The project is based on a mix of public benchmark data and proprietary simulations from industrial partners (DNV for wind turbines and Kongsberg for compressors). This will require approval from the partners before publications.
Supervisors: Alexander Stasik (SINTEF), alexander.stasik@sintef.no, with input from Andris Piebalgs (Cognite), Even Pettersen (Kongsberg), Jon Espen Ingvaldsen (NTNU)
Generating synthetic data. Banking data can contain sensitive information that limits the possibility of sharing the data openly. The data is typically restrained to being analyzed on our internal hardware. This limits the ability to use large compute clusters and cloud services. It also sets a high barrier of entry for using the data in research than what an open dataset would. If synthetic data could be created based on real data, which retains the inference and statistical relationships without containing personal data, this could make it both easier to share data and to use compute resources outside of our servers.
Thesis Description: Methods for generating synthetic data include generating data from statistical distributions, using agent based modelling, GANs and LLMs. Text data that contains personally identifiable information is hard to synthesize. Transactional data, such as card transactions can be easier to remove any personal information from, but it will probably be hard to maintain statistical relationships, especially in tandem with other customer data. One approach could be to try to synthesize an already aggregated dataset, that would contain for example the sum of transactions for a given category such as restaurants. A younger person might have a very different history of transactions. In that case, a GAN should be able to create datasets that retain the relationship between age and restaurant expenses. Another approach could be to synthesize different raw transactional data, that is connected to customer data and product data in a relational form. In that case, it would be more challenging to generate data where these statistical relationships are retained, but this would also open up a lot more use cases than synthesizing a single table or flat file.
We expect the student to synthesize banking data in such a way that the data will retain statistical information while not containing any real data. The synthesized dataset should give the same statistical inferences as the original data.
Data Description: The data would either be an already aggregated flat file, for example one of our analytical base tables for next best offer models, or a subset of relational data, such as transaction data, product data and customer data. The latter would be more challenging but also open up a lot more use cases. Students working on our data will be working on a virtual machine through a company laptop. All the data processing will have to be done on the virtual machine. The data will be de-identified.
Contact: Stian Arntsen, Stian.arntsen@smn.no
SpareBank1 SMN is the leading financial institution in middle Norway. SpareBank 1 SMN is an independent regional savings bank with a local footing and one of six owners of SpareBank1-alliansen. SpareBank 1 SMN offers competitive products in the fields of financing, savings and investment, insurance and payment services along with estate agency, leasing, accounting and capital market services. The thesis will be carried out in the advanced analytics department in the technology and development division. Relevant business resources will be included if necessary.
Supervisor (NTNU): Jon Espen Ingvaldsen, jon.espen.ingvaldsen@ntnu.no
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.
As software systems evolve, many organizations struggle with outdated documentation and legacy diagrams that describe system architectures, workflows, or business processes. These diagrams, often created using UML, flowcharts, or proprietary notations, are difficult to translate into modern programming languages. Manual conversion is time-consuming and error-prone, highlighting the need for automated solutions. Recent advancements in Generative AI (GenAI), particularly large language models (LLMs) and vision-based AI models, offer promising approaches to automate the conversion of legacy diagrams into functional code. This thesis aims to explore how GenAI can be used to interpret, analyze, and generate source code from legacy diagrams, reducing the effort required for software modernization.
Read also: Writing a Master's Thesis in Computational Creativity
To be creative, we need to produce something which is new, meaningful and has some sort of value. Generative AI is 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, or generating text based on music or images.
Domain-Specific Languages (DSLs) are tailored programming or specification languages designed for specific problem domains, such as hardware description (VHDL, Verilog), data analysis (R, SQL), and automation (BPMN, Terraform). Developing and maintaining DSLs requires domain expertise and significant effort in syntax design, compiler construction, and user documentation. Generative AI (GenAI) models, particularly large language models (LLMs) like GPT-4, have demonstrated capabilities in code generation, program synthesis, and natural language understanding. This thesis aims to explore how GenAI can be leveraged to support DSLs in various stages, including DSL creation, code synthesis, debugging, and usability improvement.
As software becomes more complex, its energy consumption and environmental impact grow significantly. The concept of green coding focuses on writing efficient, energy-saving code to reduce carbon footprints. However, optimizing code for sustainability is challenging, requiring expertise in energy-efficient algorithms, compiler optimizations, and hardware-aware programming. Generative AI (GenAI) presents a promising solution by automatically generating, refactoring, and optimizing code for better energy efficiency. AI-powered tools can suggest improvements, reduce redundancy, and enhance performance while maintaining functionality. This thesis will explore how Generative AI can assist developers in writing energy-efficient code and evaluate its effectiveness in real-world scenarios.
As digital technology advances, the environmental impact of web development has become a growing concern. Websites contribute to carbon emissions through energy-intensive processes such as server hosting, data transfer, and resource-heavy design elements. Green web design aims to reduce these environmental impacts by optimizing performance, minimizing resource usage, and improving accessibility. Generative AI (GenAI) presents a promising opportunity to enhance green web design by automating sustainable coding practices, optimizing resource efficiency, and providing AI-driven recommendations for eco-friendly development. This research explores how GenAI can assist in creating sustainable web solutions while maintaining usability and performance.
Mathematical word problems are a fundamental aspect of education, requiring both natural language understanding and problem-solving skills. Traditional methods for solving such problems rely on rule-based approaches or symbolic reasoning, but recent advances in Generative AI have opened new possibilities for automated problem-solving. Large language models (LLMs) and neural networks can now interpret, reason, and generate step-by-step solutions for complex mathematical problems.
Co-supervisor: Prof. Michail Giannakos
This master thesis focuses on leveraging Generative AI to solve visual math problems. This research aims to explore how AI can interpret and reason through mathematical problems presented in visual formats, such as graphs, geometric diagrams, or handwritten equations.
This master thesis explores the role of Generative AI (GenAI) in Software Usability Testing. This research will investigate how AI can enhance usability evaluation processes, automate testing tasks, and improve user experience (UX) assessment in software development.
A good relationship with their customers is essential for financial institutions. This involves reacting timely and adequately to customer messages. AI technology has the potential to improve customer services and become a competitive advantage for banks.
Many financial institutions rely on established machine learning models. Recently, Large Language Models (LLMs) have exceeded previous performance in many benchmarks. The candidate's task is to compare a set of LLMs to simpler models to see whether they can categorize messages more reliably.
Sparebank 1 SMN will provide a data set with more than 400,000 emails and 160,000 chats mostly in Norwegian that can be accessed with a laptop provided by them.
Stian Arntsen (Sparebank 1 SMN) will support the master project.
Naturally, organizations collect large document archives. These feature a variety of documents, some of which lack important metadata. Examples include document type, and references to products, services, or customers.
The candidate's task is to enrich the document archive by applying generative AI methods to infer missing metadata. The project requires setting up an NLP pipeline to automatically process documents into a numeric representation, and apply a set of machine learning methods to obtain estimates for missing metadata with high confidence.
The candidate will get access to a document archive from DNB and can assemble a reference data set from public sources. The project will be supported by Jan Thomas Lerstein, DNB.
The candidate will have the opportunity to engage in a summer internship at DNB in Oslo to prepare for the project. They can work on the thesis in Oslo.
Evaluate the ability of Generative AI, such as Large Language Models (LLMs), to understand and work with modeling tasks for systems and software engineering. This project can be customized in directions, such as generating code from models, generating models from natural language descriptions, modifying existing models, etc.
Semi-formal models (e.g., UML diagrams) are used in different tasks of system and software engineering, for example for documenting the system and 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. This project aims to exploit the potential of generative artificial intelligence (GenAI) to simplify and automate modeling tasks in software and systems engineering.
GenAI have seen a dramatic increase in popularity in the last few years, after the release of ChatGPT in 2022 and other generative models later. GenAI is having a major impact in many disciplines, and it is considered a major disruption also in Software Engineering and Systems Engineering tasks.
Research on the use of GenAI for software engineering tasks is emerging, for example for code refactoring tasks, writing test cases, etc. The objective of this project is to investigate the use of GenAI for modeling tasks in software and systems engineering. This project can be adapted to focus on different kind of modeling tasks and different kind of system models. Some examples include:
The long-term research objective linked to this activity is to simplify modeling tasks in software and systems engineering, through the use of GenAI.
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.
Drones are not able to navigate when GPS signals are jammed. One alternative iscelestial navigation however cloud cover makes this difficult and accurate fixes arerare. Terrain and visual features can be sensed with a camera and used to matchagainst satellite imagery or a map. However, over open water there are no real visual features as waves repeat regularly. Although, seafloor terrain and wind speed are strongly correlated with wave shape and speed which can be sensed.
The goal is to infer the seafloor terrain from waves and match again Bathymetric Maps (i.e.,seafloor terrain) in order to localize.
This project is under a collaboration with University of Waterloo, Canada and provides the possibility to take the pre-project at NTNU and the masters thesis at UofWaterloo. Taking the whole project at NTNU is also a possibility. For those that go on exchange, the canadian military is providing data. If no exchange, an agreement for this data could be reached or a different source of data applied.
A large variety of algorithms have been published to date which allow for 3D surfaces to be compared to one another. These are primarily used to detect point to point correspondences between two potentially similar surfaces. Detecting similarity is the foundation for many things you'd want to do with 3D data, such as an autonomous robot being able to recognise places it has been before.&n
These algorithms have traditionally been implemented on the CPU, but are in many cases highly parallelizable. Implementing them on the GPU is therefore a logical next step, where the higher available throughput of the processor can likely be well utilised. In turn, the faster execution times this achieves can help adopt these methods in practical environments.
The challenge is that some operations do not directly map to the GPU, and the GPU is far more memory bandwidth constrained than a CPU is. While implementing the algorithm naively is therefore likely not very difficult, adapting it to perform better on the GPU may give a number of interesting problems.
Like the goods they transport, ships will eventually become waste and need to be broken down properly. The process of ship dismantling involves various activities, and one of them is inspecting the ship to be dismantled. Such inspection is required to ensure the area to be cut does not contain materials and gases that are harmful for 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. Since ship inspectors usually do not have experience with drone operations, 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.
Like the goods they transport, ships will eventually become waste and need to be broken down properly. The process of ship dismantling involves various activities, and one of them is to cut the ship’s hulls. Currently, hulls are cut manually by workers who use scaffoldings 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 without any experience using robotic systems.
Like the goods they transport, ships will eventually become waste and need to be broken down properly. The process of ship dismantling involves various activities, and one of them is to cut the ship 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 ship 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 without any experience using robotic systems.
Automatic guitar tablature transcription is an active field in music information retrieval (MIR). It entails extracting guitar-specific music annotations from pieces of audio recordings of guitar music. Compared to other instruments such as the piano, this field is relatively underdeveloped. This is mainly due to the lack of large, high-quality datasets.Several approaches have come forward to combat this issue, but the problem remains underexplored. The main challenges this project aims to tackle are the lack of data and the exploration of transformer models utilised for automatic tablature transcription. This entails exploring brand-new datasets such as GAPS and addressing the overfitting to the GuitarSet dataset that is very prevalent in the field, as it is one of the only datasets with a sizeable amount of richly annotated guitar music recordings. Deep transformer models will be employed to transcribe pieces of guitar music. To do this, synthetic data will have to generated, as transformer models require a lot of training examples to be highly useful.
eveloping countries have limited resources for healthcare delivery hence need to make the most of resources available.
This project/ thesis is linked to the Hisp/ DHIS2 (www.dhis2.org/) initiative.
Based on open source software, Hisp aims at increasing the efficency and quality of health services by enhancing the necessary reporting of health status. Mobile technologies are crucial as, even when roads and electricity is patchy, there are mobile phones.
The approach of Hisp is pragmatic: rather than elaborate, complex 'perfect' solutions, Hisp provides simple and robust ones that have a realistic chance of uptake.
Hisp, across Africa and Asia, is implemented in about 80 countries in varying degree of completion. It is one of the world's largest systems serving patients in the Global South, measured by size of the caption population.
The project/ thesis involves empirical fieldwork in Africa or Asia on selected services of the Hisp portfolio. The purpose of the work is to identify requirements and subsequently help implement these as part of the evolving portfolio of Hisp software.
The Hisp project is managed by Univ of Oslo. This project/ thesis will be in collaboration with the UiO team.
Previous dissertations (master, Phd) and reports from the UiO archive are found here: https://www.duo.uio.no/discover (search using "DHIS2" as keyword)
The candidate(s) will preferably also be employed as part time research assistants at St Olav university hospital before or during the master project.
Project summary
Sustainable childhood obesity prevention requires common knowledge, planning and actions from municipality, healthcare, schools, community organisations, parents and children.
In Norway, we may combine comprehensive anonymized demographic, school and health data for basic statistical geographical units covering 200-2000 persons (SSB: Grunnkrets). We know that some neighbourhoods have good overall child health, facilities and environments fostering activity and preventing obesity, school participation, while other areas are facing challenges. Being able to see, monitor, intervene and design for better child health.
The project’s aim is to develop one or more prototypes of decision support and data visualization applications in this area, using the data mentioned above. This development will be done in close collaboration with important stakeholders in St. Olav and Trondheim municipality, as well as representatives from schools, parent organizations, commercial developers and food chains.
Referanser:
Supervisors:
A digital twin is defined as a virtual representation of a physical asset, or a process enabled through data and simulators for real-time prediction, optimization, monitoring, control, and informed decision-making. This project collaborates with prof. Adil Rasheed from Institute of cybernetics. Example of digital twins relevant to this project are: an autonomous aquarium or greenhouse, an experiment of soil movement and an experiment of overload prevention in electric cables. The master thesis will focus on the development of an VR environment for visualizing of and interacting with the digital twin.
A large number of algorithms have been developed over the years intended to do 3D shape matching. The recent interest surrounding machine learning has only accelerated it. However, the main focus thus far has been on recognising only the shape of an object. Using colour information as well for this purpose has received very little attention.
What complicates matters further is that while it's possible for anyone to come in and come up with matching algorithms that use colour information, there would have to be some kind of way to evaluate how well such an algorithm is working. As of right now, no such benchmark exists.
The appearance (colour) of an object can be affected by many things. The colour of the light that is incident on the surface, the shininess of the material, and its translucency are only a few examples.
We have previously developed a benchmarking strategy that evaluates how much a given matching algorithm is affected by variations in shape. This project aims to extend this benchmark to also evaluate how different factors influencing the colour or appearance of an object can affect its matching performance. The colour lab in Gjøvik can also assist in this project with their expertise.
3D object classification is the process of taking a 3D object (such as an airplane or a chair), and using its 3D surfaces to determine which of a limited set of classes that object belongs to. The methods that at this time perform best at this are deep learning based approaches. It's also a fairly hot topic, with dozens of papers having been published in recent years, including at top tier AI conferences.
The de facto benchmarks to test these methods are Modelnet40 and ScanObjectNN. These respectively contain 12,311 (9,843 training, 2,468 testing) and ~15,000 objects. While not an insignificant amount of data, the margins between the improvements added by each paper is often only in the order of 0.1%, or a handful of objects more or less being identified correctly.
It is therefore relevant to test these methods on larger benchmarks, which are now available. This hopefully allows us to gain insight in how well each of these methods perform in relation to each other on more quantitative “solid ground”.
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.
Contemporary works on Human-Centered AI (HCAI) focus on creating AI systems that amplify and augment rather than displace human abilities. HCAI seeks to preserve human control in a way that ensures AI meets our needs while also operating transparently, delivering equitable outcomes, and respecting privacy. AI systems function in diverse spaces (e.g., social, work, and classroom) alongside traditional interactions and activities. Therefore, it is expected that humans and AI will complement each other, stand by each other, and engage in a process of co-learning, co-creation, and co-evolution. Such a process is necessary for combining the strengths of humans and AI and reinforcing each other to achieve Hybrid Intelligence (HI). Unlike traditional AI, designed to operate independently in performing tasks that typically require human intelligence, such as perception and learning, HI involves active collaboration between humans and machines. Thus, further work is needed to understand and design appropriate HCAI technology, with a particular focus on how teachers can work together with AI tools to synergistically combine their strengths to reinforce efficient, and ethical use of technology.
In this topic, the augmentation perspective and the concept of HI will be used to guide this work. The candidates will engage with the design (co-design or participatory design) of learning services (e.g., interfaces or other artefacts) to showcase the challenges and opportunities of hybrid human‐AI learning technologies. The six levels of automation model will be used to identify the roles of the various AI users (e.g., learners, teachers). The transition of control between teacher or students and technology needs to be articulated at different levels and related to the augmentation perspective.
For this topic, there is an option to collaborate with an EdTech company called LearnLab (see: https://www.learnlab.net/en/). LearnLab's platform includes innovative web applications like Colab, Storylab, Idealab, Medialab, and Mylab. These tools support everything from interactive teaching and brainstorming to multimodal storytelling and the production of videos, podcasts, and formative assessments. Through Learnlab’s learner-focused AI, both teachers and students receive personalized support and formative feedback, with the goal of enhancing the learning outcomes and saving teachers' time.
The nature of decision-making is changing drastically, both in personal lives and in the business sphere. An increasing amount of decisions are now based on insight that is generated through analytics. Despite this, often individuals are faced with cognitive-overload, conflicted views, or biases that result in non-adoption of insight. This project will be done in collaboration with the Big Data Observatory (https://www.observatory.no) and involve designing a study protocol and collecting and analyzing neurophysiological data (eye-tracking and electroencephalography) from study participants. This will be done with the help of an expert in such tools.
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 hatespeech 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 various machine learning methods to identify such language.
Recognizing and treating sepsis is challenging because symptoms overlap with other diseases and the patient population is diverse. To improve treatments and understand different patient groups, many studies focus on identifying different sepsis patient groups using clinical phenotypes. Clinical phenotypes include a patient’s signs, symptoms, conditions, and in-hospital events.
The objective of the master’s thesis is to utilize structured and textual data to identify sepsis clinical phenotypes for different patient groups. Work will be performed on medical records from St. Olavs Hospital. This project is a step towards identifying patients at risk for sepsis.
Number of students: 1-2
eParticipation (electronic participation) is defined as the use of ICT to facilitate citizen involvement in decision-making processes. This approach is receiving increasing importance in urban planning, where inhabitants are invited to share their ideas and co-design urban environments. The immersive technologies of AR and VR present new opportunities for these processes through enhanced visualization, communication, and informed decision-making. The master thesis will focus on the development and evaluation of immersive environments (AR or VR) for the visualization of urban design solutions and the facilitation of participatory practices. The thesis can be carried out in collaboration with Enact15mc project, focusing on the co-design of HaakonVII gate in Trondheim.
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.
More specifically, the goal of the project is the following:
[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.
The aim of the project is to implement and evaluate a scalar evolution analysis 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 [6]. In contrast, the Regionalized Value State Dependence Graph (RVSDG) isa 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.
Scalar Evolution is a compiler analysis that looks at the change in the value of scalar variables over iterations of a loop. The analysis provides facts about induction variables that are utilized in other loop transformations and simplifications, such as loop strength reduction or loop invariant code motion, to improve the quality of the loop code based on these facts.
Currently, the RVSDG is implemented in the JLM compiler [1]. The aim of this project is to add a scalar evolution analysis to JLM and evaluate the implementation utilizing loop simplifications and transformations. 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:
[2] Olaf Bachmann, Paul S. Wang, and Eugene V. Zima. Chains of recurrences—a method to expedite the evaluation of closed-form functions. In Proceedings of the International Symposium on Symbolic and Algebraic Computation, 1994. [Online]. Available: https://doi.org/10.1145/190347.190423
[3] Johnnie L Birch. Using the chains of recurrences algebra for data dependence testing and induction variable substitution. PhD thesis, Florida State University, 2002.
[4] Robert Engelen. Symbolic evaluation of chains of recurrences for loop optimization. 2000.
[5] Robert van Engelen. Efficient symbolic analysis for optimizing compilers. In Proceedings of the International Conference on Compiler Construction, 2001.
[6] 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. [Online]. Available: https://doi.org/10.1145/3391902
[7] Eugene V. Zima. On computational properties of chains of recurrences. In Proceedings of the 2001 International Symposium on Symbolic and Algebraic Computation, 2001. Online]. Available:https://doi.org/10.1145/384101.384148
Health literacy—the ability to access, understand, and apply health information—is crucial for making informed health decisions. However, many individuals struggle with low health literacy, leading to poor health outcomes. Traditional health education methods often fail to engage audiences effectively. Gamification, the use of game design elements in non-game contexts, has emerged as a promising strategy to enhance learning and engagement. This research aims to explore how gamification can be effectively integrated into health education to improve health literacy. The study will focus on designing and evaluating a gamified learning system that encourages users to acquire, retain, and apply health-related knowledge.
While conversational AI and even image & video analysis enjoy widespread use, generative 3D is still nascent, and models and approaches are more experimental. This thesis will investigate different approaches both in-memory and hosted via API use regarding their applicability to support scene composition for a variety of educational scenarios. Aim is to create and evaluate a proof-of-concept in combination of Python back-end and Unity3D front-end app, integrated into the existing source-code projects MirageXR (Unity3D) and lxr (Python, Django) to benefit from existing development.
With this master thesis project, you will:* Design and develop an architecture for generating 3D models from prompts and managing API communication for submitting jobs and retrieving generated 3D models* Assess its feasibility and evaluate efficacy with a small-scale user experiment* Investigate potential educational usage scenarios (e.g. medical simulation, language learning, creative writing)
Outline solution:* submit prompt to web service to initiate 3D generation* monitor whether generation process has finished* download 3D artefact and display in MirageXR
Additional informationAim is to interface this service with MirageXR (https://github.com/WEKIT-ECS/MIRAGE-XR/), the AR learning experience editor and player, to support creation of 3D learning content based on user description. This can be used to, e.g., create props and objects required in XR learning activities.
How to:* Here is an example API: https://api-documentation.blockadelabs.com/api/* Here is an alternative (using single-shot image input): https://github.com/VAST-AI-Research/TripoSR* And the Shap-E implementation: https://github.com/openai/shap-e/tree/main?tab=readme-ov-file* And 4dfy: https://sherwinbahmani.github.io/4dfy/ and https://sherwinbahmani.github.io/4dfy/
ContextThe students will have access to a very well-equipped IMTEL VR lab (https://www.ntnu.edu/imtel/) containing various modern AR and VR devices and laptops.The target devices for the project are Apple Vision Pro, HoloLens 2, and Oculus Quests 3.
Main contactFor any questions about the task, please, contact Mikhail Fominykh mikhail.fominykh@ntnu.no.
SupervisorsProf Dr Monica Divitini, Professor at the Department of Computer Science, NTNUDr Mikhail Fominykh, Researcher at the Department of Education and Lifelong Learning, NTNUProf Dr Fridolin Wild, Professor AR/VR at the Open University, United Kingdom
Emerging technologies such as virtual/augmented reality/extended reality (VR/AR/XR) and generative AI such as ChatGPT, Midjourney and Magic3D are already revolutionizing how we live and work. XR has already demonstrated significant potential in transforming educational practices by providing learners with realistic and highly engaging learning experiences. Generative AI is a powerful tool that can be used to quickly and efficiently create a wide range of educational content, including human-like text, videos, images, and even 3D models and software code. The goal of this master project to investigate if the combination of these technologies can contribute to creating innovative education tools for NTNU teachers and students.
There are 2 possible research questions in this project:1. Development and evaluation of virtual classrooms and learning areas, populated by virtual humans/ teaching assistants powered by ChatGPT or similar chatbots (in collaboration with NTNU teachers, using existing ChatGPT plugins). These teaching assistants will be able to interact with students 24/7, answering their questions and providing assistance and personalized feedback. 2. Explore how generative AI (Magic3D or similar) can be used to rapidly create 3D educational content for use in such virtual classrooms by teachers/students without prior programming experience. Previous studies among teachers at NTNU and similar international studies identified difficulties with generation of educational content as one of the major obstacles to wider adoption of XR among educators.
The students will have access to a very well-equipped IMTEL VR lab (https://www.ntnu.edu/imtel/) containing 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, VR laptops etc. A significant number of the VR/AR equipment is portable and can be used at home shall the pandemic situation and campus closure be repeated.
Supervisors: Monica Divitini, Ekaterina Prasolova-Førland (ekaterip@ntnu.no) & Mikhail Fominykh!!!PLEASE CONTACT Prof. Prasolova-Førland for more information about the task!!!
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:
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.
The job market, in Norway and internationally, has changed considerably over the past few years due to the COVID-19 pandemic and the emerging AI technologies, raising the need for developing innovative methods for workplace training and career guidance. In this project we will investigate how the use of Virtual Reality technologies and gaming elements can 1) motivate and inform young job seekers on their way to work and 2) contribute to faster skill acquisition for new employees. Through the simulation of a workplace or an industry (e.g. aquaculture or a shipyard), the job seekers can immerse into different workplaces and try out typical tasks, for example, salmon feeding or welding in a safe setting, thus mastering the corresponding real world situation.The master project will be performed in collaboration with Erasmus+ VR4VET project (Virtual Reality for Vocational Education and Training, https://vr4vet.eu/) involving several partners in Norway, Germany and Netherlands. The project proposes a new approach to vocational training and career guidance applying VR to allow active and engaging exploration of professions and introductory training, involving job seekers, career counsellors and industry stakeholders all over Europe. The student(s) will work in close collaboration with NAV, local industries (especially maritime) and our European partners (TU Delft and TH Köln). VR4VET is a continuation of Virtual Internship project that has so far resulted in several prototypes for workplace training and job interview training in VR and received international recognition (e.g. Best Demo Award at EuroVR 2018 and Breakthrough Auggie Award finalist) and broad media coverage https://memu.no/artikler/gir-ungdom-en-virtuell-jobbsmak/, https://www.ntnu.edu/imtel/virtual-internship.
NB: for many of the industri proposals listed below a research assistant position can be offered in additon, in some cases even a summer job.
As some if you have reported issues with the direct visual intelligence related links provided below I also include the link to all the projects pitched thought the AI lab here.
MIA (Monitoring-Insight-Action): Your data. Your digital twin. Your health (video)
MIA (Monitoring-Insight-Action): Your data. Your digital twin. Your health (text)
Rosenborg (RBK) - Analyseavd.:
Sport Analytics: Fotball (RBK) - AI and Computer Vision for video analysis (pres)
Sport Analytics: Fotball (RBK) - AI and Computer Vision for video analysis (text)
Jotun:
A Cutting Edge Tool for Analyzing Fouling on Test Panels (video)
A Cutting Edge Tool for Analyzing Fouling on Test Panels (pres)
A Cutting Edge Tool for Analyzing Fouling on Test Panels (text)
Hydro:
Use of AI for Grain Size Characterization of Aluminium Alloys (video)
Use of AI for Grain Size Characterization of Aluminium Alloys (pres)
Use of AI for Grain Size Characterization of Aluminium Alloys (text)
St.Olavs / NTNU:
AI-assisted labeling of coronary arteries from CT-data: A next step from invasive to non-invasive diagnosis of coronary artery disease (video)
AI-assisted labeling of coronary arteries from CT-data: A next step from invasive to non-invasive diagnosis of coronary artery disease (pres)
AI-assisted labeling of coronary arteries from CT-data: A next step from invasive to non-invasive diagnosis of coronary artery disease (text)
SINTEF / SVV:
Autonomous Driving (AD) under Nordic conditions (video)
Navigation of automated vehicles in nordic conditions 1: Develop AI / CV methos for generation og HD-maps / Digital Road Twin (text)
Navigation of automated vehicles in Nordic conditions 2: Identify, test and evaluate AI-algorithms for navigation of AVs with LiDAR and/or video data (text)
Aerial laser classification (video)
Aerial Lidar segmentation using deep learning (text)
Aerial Batymetric Lidar, classification using full waveform and deep learning (text)
Proxpect Drones:
DEIDAC: Drone External Inspection Detection using Algorithms and Computer Vision (video)
DEIDAC: Drone External Inspection Detection using Algorithms and Computer Vision (pres)
DEIDAC: Drone External Inspection Detection using Algorithms and Computer Vision (text)
AI, AI, AI, what a gentle lobster:Huge potensial for Norwegian food industri (lobster forming) and avery interesting CV/AI project, involving detection and tracking (objects and key points), as well as the final classification as a function of time (green=gentle, yellow=undecided, red=not gentle). We have large amounts of data, incuding high quality labelded data.
Pres
Text
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.
This master assignment will apply Large Language Models (LLM) in the analysis and documentation of Systems Engineering tasks for developing Air Traffic Management Systems (ATMS). It is relevant to air traffic control and technology development in multiple countries. The project is in collaboration with Avinor, within the iTEC SkyNex context.
Your benefit – Artificial Intelligence and/or Systems Engineering in a European project
Would you like your master thesis to be relevant to air traffic control and technology development in United Kingdom, Netherlands, Germany, Spain, Canada, Poland, Lithuania and Norway? With this master assignment you will take part in the largest software development project for ATMS in the world. Your contacts network can expand to Europe and Canada, and open career opportunities in Norway, Canada and Europe.
Project background – ATMS and the Systems Engineering Complex – iTECSkyNex.com
When developing an ATMS, the Systems Engineering is one of the major efforts and costs. About 80% of the Systems Engineering involves analysing and documenting the system, whereas the remaining 20% is the implementation, typical coding the software. Whereas much effort is made to investigate and use artificial intelligence for generating code these days, less time is spent on the 80% of Systems Engineering for analysing and documenting systems.
When analysing and documenting an ATMS, domain knowledge needs to be elicited, analysed and understood. Many stakeholders are involved in these tasks and domain knowledge needs to be transformed and understood by typical domain and subject matter experts, requirements, system and software engineers, to develop the system. The tasks of the engineers involve creating specifications, from the overall system descriptions, breaking them down into system architecture and design descriptions, and further down into descriptions of the software (also architecture and design). As specifications must be understood by many stakeholders, the use of natural language requirements (NLR) in combination with diagrams at various abstraction levels is common. Decomposing and tracing the information into more details is a key activity, together with the verification and validation of the information and the ATMS being developed.
Structure of master assignment
Main research questions:
Methods:
Delivery and expected results:
Specific knowledge/competence/skills for this assignment:
This project will be co-supervised by Leonardo Montecchi and Jingyue Li
In this project, the joint interests of advisor and student(s) come together.
Typically, the project will be based on previous research by Prof. Ole Jakob Mengshoel AND interests of students. If only one of these is present, there is no basis for a project.
Please send email to me if you're interested. I will only accept students for the project based on mutual understanding of such a project.
As educational institutions adopt various digital learning platforms, seamless interoperability and data integration become essential for enabling effective learning analytics (LA). This thesis explores interoperability frameworks such as the Experience API (xAPI) and Learning Tools Interoperability (LTI) within Norway’s learning ecosystem, explicitly focusing on systems like FS, Canvas, and other data flows managed by Sikt.
The aim is to develop a solution that accesses and aggregates learning data across these systems via available APIs, offering stakeholders meaningful insights through a dashboard or analytics tool.
Thesis Description
The thesis starts with a literature review covering learning analytics, xAPI, LTI standards, and educational data interoperability. Students will then design and implement a prototype system capable of:
The system will serve teachers, students, and administrators, providing a unified view of learning progress, resource usage, and interaction patterns.
Candidates should be comfortable with software development and interested in educational technology and data science. Required skills include:
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.
Let's say we have a robot roaming around a building. While walking around, it creates a 3D capture of its surroundings. This capture can subsequently be used to more accurately determine its position. Various kinds of sensors that sense motion have a tendency to drift over time, so using the surroundings for navigation helps to correct where the robot believes it is currently located. This process is known as Simulatenous Localisation and Mapping (SLAM).
One way in which the robot can correct its location is to recognise a place it has seen before. It therefore takes any new 3D surfaces it sees, and goes back into its previous 3D scans to try and see if it recognises what it is looking at now from an earlier point in time. If there is a clear match, it can retroactively correct the path it believes it has walked, as well as reduce any errors in the 3D scan itself.
The process of determining whether a given surface has been seen before requires an efficient 3D matching algorithm. The thing is, the more data that algorithm has to process, the longer that process takes. One manner by which previous literature has solved this is to compute so-called “keypoints” in the environment; points of interest that are likely going to be unique, and discarding all the others as they are likely to not be of interest.
The key idea of this project is that the keypoint solution does not take into another source of information that is sitting in plain sight: the 3D data we have seen thus far. Keypoint detectors usually do a good job of filtering away “boring”/unlikely to be unique surfaces like flat walls. However, if all offices in a building have the same chairs in them, answering the question of which office you're in when you see such a chair is inconclusive, because that chair is not very unique in relation to the 3D surfaces you have seen.
This project aims to create a keypoint detector that takes an index tree of 3D surface patches represented as so-called local 3D shape descriptors, and as new surface patches are added to this index, it evicts any surfaces that it deems as non-unique from the index. What remains are surfaces that are unique with respect to the dataset as a whole.
We can take this project into two primary directions. First, we can look at what unique surfaces look like in the context of a large dataset. And second, it would be possible to see if a methodology like this allows unique locations in a 3D scan of a building to be located in realtime.
Agile software development methods have gained increasing popularity in software development projecs. Agile methods prescribe practices for development, and were first used in small projects with little criticality. However, such methods are increasingly used in large projects, and this project will investigate how the practices are adapted and combined with traditional practices to function effectively in large scale. 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 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
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.
Supervisors: Michail GiannakosPlace: LCI Lab: https://lci.idi.ntnu.no/Suitable for: One or two students
IntroductionLearning analytics (LA) and AI in education (AIED) have been hot topics in educational communities, organizations and institutions. There are four essential elements involved in all LA and AIED processes: data, analysis, report and action.
Learning analytics are important because every “trace” within an electronic learning environment may be valuable information that can be tracked, analyzed, and combined with external learner data; every simple or more complex action within such environments provides insights that can guide decision-making (e.g.,, students, teachers, policymakers).
Thesis DescriptionThe increased need to inform decisions and take actions based on data, points out the significance of understanding and adopting LA and AIED in everyday educational practice. To treat educational data in a respectful and protected manner, the policies for LA play a major role and need to be explicitly clarified. This thesis will analyse data associated with the use of LA and AI learning systems in Norway, and has the option to also collect primary data (e.g., questionnaires or interviews with students and lecturers), with an ultimate goal to identify what LA and AIED systems are used in Norway, how they are put into practice and potential challenges and opportunities with their us.
RequirementsThe ideal candidate will have a background and interest in data analysis and research methods, no programming skills are required.
Relevant informationThe ideal candidate will have a background in data analysis, no programming skills are required.- The candidate can use data that have been collected in the context of the national expert group in learning analytics: https://laringsanalyse.no/- The candidate can use data from different national organizations such as Sikt and NOKUT.
See the complete topic as PDF: https://drive.google.com/file/d/1wv5l2eok3LLfTuGucurgHEJ7Z9RSztVc/view?usp=sharing
How do software companies ensure organisational learning at different organisatinoal levels? Learning organisations and knowledge management have been vital areas for software organisations. Recently, many organisations have focused their learning activities at team level rather than individual level. This project will first survey literature on organisational learning and knowledge management in software engineering, and as a possible extention in a master thesis conduct an empirical study in a Norwegian software engineering environment. The student can suggest a case, or a case can be found through the supervisor´s network.
Deployment of new technological infrastructures such as platforms and AI requires new skills to be learned. However, it is not easy for busy practitioners to attend classes as students do. Many people use online resources such as YouTube and social media to keep updated, but this learning is seldom done systematically. We need new pedagogical models to keep updated on the job. We want you to find and design learning models for busy people who need to keep their skills updated all the time.
This is an empirical research projects. You will need to create empirical knowledge about new learning models through methods such as co-design and case studies. Your results will include new knowledge, but also models and design ideas for new learning models and tools.
This task requires that you have a good understanding of, and are interested, in empirical qualitative research. Working language for this task is English. The thesis can be written in Norwegian or English but we recommend English. Please contact Babak before you select this task.
This project proposes a Large Language Model (LLM)-based multi-agent system to enhance the development and operation of embedded systems in IoT environments. In this architecture, each autonomous agent—powered by an LLM—controls a specific IoT element such as lighting, motors, sensors, or actuators, and communicates with other agents to coordinate system-wide behaviors.
The agents operate using natural language prompts and context-aware reasoning to interpret tasks, optimize device performance, and adapt to dynamic conditions. For example, a sensor agent can relay data patterns to a lighting agent to adjust brightness in real-time, or a motor control agent can respond to both environmental inputs and user commands, all without hardcoded rule sets.
This approach enables a more modular, intelligent, and adaptable embedded system, where new devices can be integrated simply by assigning a capable agent and defining its high-level goals. By combining LLM reasoning with distributed IoT control, the project aims to reduce development time, improve system flexibility, and support complex decision-making at the edge.
Large Language Models (LLMs) are powerful tools for tasks in Natural Language Processing (NLP). They take in vast amounts of texts and learn statistical patterns. Still, domain-specific language resources are required such as efforts in fine-tuning, continued pre-training, or Retrieval Augmented Generation (RAG) to support use cases.
The candidate will obtain a set of documents from the maritime domain and examine the utility of different LLMs and their augmentations for use cases related to maritime language. For instance, the candidate can use maritime regulations to optimize a set of LLMs to answer questions related to maritime regulations.
The project will be supported an co-supervised by Børge Rokseth (NTNU), and Rudolf Mester (NTNU).
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 goal of this project is to apply techniques from machine learning to optimize the design of nanomagnetic computing systems. This is an interdisciplinary project at the intersection between computing, artificial intelligence and physics.
A key challenge in ASI computing systems is optimizing the placement and orientation of the magnets such that some desired functionality is obtained. With thousands of magnets, the parameter space is way too big for exhaustive search methods. However, modern machine learning methods are routinely applied to optimize neural networks with billions of parameters. Frameworks such as PyTorch makes these methods readily applicable to non-neural systems. For instance, PyTorch was recently employed to optimize a nanomagnet-based spin-wave scatterer to perform computing tasks such as vowel recognition. This was made possible by incorporating PyTorch gradients into the physical simulator itself.
Within COMET we are exploring computing in ASI systems. Towards that end we have developed a large-scale ASI simulator called flatspin. Integrating PyTorch (or similar) into flatspin should open for the possibility to optimize ASI designs towards some target function. Alternatively, other machine methods such as evolutionary search may also be relevant.
Artificial Spin Ice (ASI) are metamaterial systems consisting of coupled nano magnets arranged on a 2D lattice, whose collective large-scale emergent behavior has attracted considerable interest as computing substrates.
The project will be carried out in collaboration with the Norwegian Institute for Nature Research (NINA). 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. You will be collaborating with the Miljødata department with researchers who focus on employing advanced technologies to study and protect biodiversity, with a particular emphasis on bioacoustics and rare species.
Two datasets are available: One dataset with arctic fox vocalization and one dataset with Norwegian bird song (aka the Sound of Norway (see https://thesoundofnorway.com/) dataset. The main themes of the proposed project would remain the same regardless of the dataset chosen, but the rest of the proposed project is written with the Sound of Norway dataset in mind. It is up to the MS students which dataset they want to start working on.
We have, for the Sound of Norway dataset, used an off-the-shelf deep learning-based bird classifier (i.e. BirdNET) that effectively identifies common bird species. However, its performance significantly drops for the rarer species, limiting our understanding of biodiversity in diverse ecosystems. Properly accounting for rare species is critical to establish ecosystem health and it is essential to develop automatic methods efficient at detecting and classifying them.
The project aims to (i) refine the model using advanced machine learning techniques (such as fine-tuning and few-shot learning strategies) and (ii) explore other innovative AI methods to enhance species recognition. This is particularly important for underrepresented species. The expected outcome includes improved model accuracy and recall for rare species, contributing to more comprehensive biodiversity assessments.
Data: The student(s) will have access to the Sound of Norway dataset, which contains approximatively 7 terabyte of audio files recorded by BUGG recording devices at 41 different sites throughout Norway. All of the detections made by BirdNET have been verified by an ornithologist. Moreover, a few files have completely been reviewed by the ornithologist to compute the model recall. This dataset should provide a robust foundation for the evaluation of methods you will be developing.
This thesis explores the development of a machine learning model that serves as a real-time pace indicator for Formula Student endurance racing. Using sensor data, the model predicts optimal driving speeds that balance competitiveness with reliability, supporting engineers and drivers in maximizing race performance and finishing probability.
Problem Statement: Many organizations have been caught up in the tide of the AI-promises. Yet, we see many examples of a specific problem not being solved or the gap between capabilities and intentions without a bridge. Leading to a misalignment between what AI can actually deliver, and the goals organizations hope to achieve. In this project, we seek to study areas for AI use cases in the municipality. It is possible that taking a responsible AI perspective will enable aspects such as the ethical, explainable, and transparent for implementation to be considered as well.
The thesis can involve:
OsloKommune Context: 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: Casandra Grundstrom (sociotechnical), Elena Parmiggiani (sociotechnical) as co-supervisor in 25/26
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.
Co-supervisor: George Adrian Stoica
Project Background
This project is in the area of artificial intelligence for stochastic optimization. The emphasis will be on the integration of methods from evolutionary algorithms and stochastic local search, with a focus on both theory and applications.
In artificial intelligence and machine learning, there are several methods that use or rely on randomization: mutation and crossover in evolutionary algorithms [Goldberg, 1989], probabilistic acceptance in simulated annealing, dropout and stochastic gradient descent in deep learning, stochastic local search (SLS) [Selman et al., 1992] [Hoos & Stützle, 2005] [Mengshoel et al., 1998], and randomization in systematic search [Gomes et al., 1998].
Stochastic optimization (SO) algorithms, which include stochastic local search (SLS) and evolutionary algorithms (EAs), will be a main focus in this project. Stochastic optimization (SO) algorithms are among the best in solving computationally hard problems including NP-hard problems such as satisfiability (SAT), traveling salesman (TSP), scheduling, and most probable explanations in Bayesian networks (BNs) [Selman et al., 1992] [Hoos & Stützle, 2005] [Mengshoel et al., 2010] [Mengshoel et al., 2016] [Pal & Mengshoel, 2016].
Research on SO performed by me and my collaborators has focused on (i) improving the theoretical foundation of SO; (ii) empirically demonstrating the benefit of SO using natural and synthetic problem instances; and (iiii) applying the approach in diverse areas of artificial intelligence and machine learning. For (i), we have for instance analyzed homogenous Markov chains to derive instructive expected hitting time results [Mengshoel, 2008] [Mengshoel et al., 2016]. For (ii) and (iii), we have demonstrated that SO algorithms can be competitive for applications including computing the most probably explanation (MPE) in Bayesian networks; feature selection in machine learning; and sparse signal recovery in signal processing [Mengshoel et al., 2010] [Mengshoel et al., 2011] [Mengshoel et al., 2014] [Pal & Mengshoel, 2016] [Yu et al., 2017] [Liu et al., 2018].
Despite several successes, the current state-of-the-art suffers from some drawbacks, creating research opportunities in several areas. In this research project, we will explore those opportunities and develop novel ideas at the interface of stochastic local search, machine learning, evolutionary algorithms, probabilistic graphical models, and deep learning. Special attention will be paid to multimodal functions, which play a central role in artificial intelligence and evolutionary algorithms. There appears to be several limiting factors when it comes to existing work on SO for such multimodal functions, including the following. Firsrt, multimodal optimization algorithms often search for “as many local optima as possible” without input or guidance from a decision maker. Second, evolutionary algorithms have been used for certain combinatorial optimization problems including feature selection, but there has been surprisingly little emphasis on diversity maintenance including niching in research on machine learning including feature selection.
The integration of local search and evolutionary algorithms has, to some extent, been studied in the literature, often under the term memetic algorithms (MAs) [Eiben & Smith, 2015] [Krasnogor & Smith, 2005] [Xue et al., 2016]. Intuitively, the goal of an MA is for the EA component to search well globally, while the SLS component’s goal is to optimize locally. In other words, one wants “the best of both worlds.” There is a broad diversity of different MA types and designs [Krasnogor & Smith, 2005] [Xue et al., 2016]. Both Lamarckian and Baldwinian MAs have been studied, reflecting the spirited discussion between Lamarck and Baldwin in evolutionary biology in the late 1800. Current research seems to favor Lamarckian MAs, in which the locally optimized individual replaces the original individual in the population.
Focus in Project
The focus of the project is on the integration of SLS and EA methods, inspired by so-called memetic algorithms (MAs), to be analyzed and tested on synthetic and natural problems. Several MA-based approaches to feature selection [Kannan & Ramaraj, 2010] [Xue et al., 2016] [Lee et al., 2019] have been developed, inspiring this proposed research project.
To be investigated are methods that addresses the issues identified above, and perhaps others, by integrating multiple different methods, and carefully balancing their execution via parameter tuning and control [Krafotias et al., 2015] [Mengshoel et al., 2014] [Mengshoel et al., 2016] [Yu et al., 2017]. An interesting idea is to combine stochastic local search, evolutionary computing, clustering, and feedback control algorithms, with emphasis on the balance between exploration and exploitation during search. Such balance can be achieved by means of niching, for example crowding [Mengshoel et al., 2014]. The goal will be to empirically test such a multi-method algorithm on synthetic and natural combinatorial function optimization problems, including ML problems (for example feature selection or network search problems), and empirically establish performance relative to other methods. For the more theoretically oriented student, theory development would be of great interest [Mengshoel, 2008] [Nguyen & Sudholt, 2020].
The project can be suitable for one or two students. The students will participate in one or more of the following tasks in this proposed project:
We already have implementations of MA, SLS, and EA methods that can be used as starting points in the project. We also have concrete ideas about problems and datasets to be considered, based on the interest of the student or students.
[Eiben & Smith, 2015] A. E. Eiben and J. E. Smith. 2015. Introduction to Evolutionary Computing (2nd ed.). Springer-Verlag Berlin Heidelberg, Chapter 9.1.
[Elsken et al., 2019] T. Elsken, J. H. Metzen, and F. Hutter. Neural Architecture Search: A Survey. Journal of Machine Learning Research, Vol. 20, Iss. 55, 2019.
[Goldberg, 1989] D. E. Goldberg. Genetic Algorithms in Search, Optimization and Machine Learning (1st ed.). Addison-Wesley, Boston, MA, 1989.
[Gomes et al., 1998] C.P. Gomes, B. Selman, and H. Kautz, Boosting combinatorial search through randomization, in: Proc. of the Fifteenth National Conference on Artificial Intelligence (AAAI-98), Madison, WI, 1998, pp. 431–437.
[Hoos & Stützle, 2005] H.H. Hoos and T. Stützle, Stochastic Local Search: Foundations and Applications, Morgan Kaufmann, San Francisco, CA, 2005.
[Kannan & Ramaraj, 2010] S. S. Kannan and N. Ramaraj. A novel hybrid feature selection via Symmetrical Uncertainty ranking based local memetic search algorithm. Knowledge-Based Systems 23, 6 (2010), pp. 580–585. https://doi.org/10.1016/j.knosys.2010.03.016
[Krafotias et al., 2015] G. Krafotias, M. Hoogendoorn, and A. E. Eiben. 2015. Parameter Control in Evolutionary Algorithms: Trends and Challenges. IEEE Transactions on Evolutionary Computation 19, 2 (2015), pp. 167–187.
[Krasnogor & Smith, 2005] N. Krasnogor and J. Smith. 2005. A tutorial for competent memetic algorithms: model, taxonomy, and design issues. IEEE Transactions on Evolutionary Computation 9, 5 (2005), pp. 474–488.
[Lee et al., 2019] [24] J. Lee, I. Yu, J. Park, and D.-W. Kim. Memetic feature selection for multilabel text categorization using label frequency difference. Information Sciences 485 (2019), pp. 263 – 280.
[Liu et al., 2018] B. Liu, T. Yu, I. Lane, and O. J. Mengshoel. Customized Nonlinear Bandits for Online Response Selection in Neural Conversation Models, in Proc. AAAI 2018.
[Mengshoel, 2008] O. J. Mengshoel. Understanding the Role of Noise in Stochastic Local Search: Analysis and Experiments. Artificial Intelligence Vol. 172, Iss. 8-9, 2008.
[Mengshoel et al., 2010] O. J. Mengshoel, D. Roth, and D. C. Wilkins. Portfolios in Stochastic Local Search: Efficiently Computing Most Probable Explanations in Bayesian Networks. J. Autom. Reasoning, 2010.
[Mengshoel et al., 2011] O. J. Mengshoel, D. C. Wilkins, and D. Roth. Initialization and Restart in Stochastic Local Search: Computing a Most Probable Explanation in Bayesian Networks. IEEE Transactions on Knowledge and Data Engineering, Vol. 23, Iss. 2, 2011.
[Mengshoel et al., 2014] O. J. Mengshoel, S. F. Galán, and A. De Dios. 2014. Adaptive generalized crowding for genetic algorithms. Inf. Sci. 258 (February, 2014), pp. 140–159. DOI: https://doi.org/10.1016/j.ins.2013.08.056.
[Mengshoel et al., 2016] O. J. Mengshoel, Y. Ahres, and T. Yu. Markov Chain Analysis of Noise and Restart in Stochastic Local Search, in: Proc. of the Twenty-Fifth International Joint Conference on Artificial Intelligence (IJCAI-16), 2016.
[Nguyen & Sudholt, 2020] P. T. H. Nguyen and D. Sudholt. Memetic algorithms outperform evolutionary algorithms in multimodal optimisation, Artificial Intelligence, Volume 287 (2020), pp. 1 - 21.
[Pal & Mengshoel, 2016] D. K. Pal and O. J. Mengshoel. Stochastic CoSaMP: Randomizing Greedy Pursuit for Sparse Signal Recovery, in: Proc. European Conference on Machine Learning and Principles and Practice of Knowledge Discovery, p. 761 – 776, 2016.
[Selman et al., 1992] B. Selman, H. Levesque, and D. Mitchell. A new method for solving hard satisfiability problems, in: Proc. of the Tenth National Conference on Artificial Intelligence (AAAI-92), San Jose, CA, pp. 440–446, 1992.
[Xue et al., 2016] [45] B. Xue, M. Zhang, W. N. Browne, and X. Yao. 2016. A Survey on Evolutionary Computation Approaches to Feature Selection. IEEE Transactions on Evolutionary Computation 20, 4 (2016), pp. 606–626.
[Yu et al., 2017] T. Yu, B. Kveton, and O. J. Mengshoel. Thompson Sampling for Optimizing Stochastic Local Search, in: Proc. ECML PKDD, 2017.
Mental health and wellbeing are increasingly recognized as critical factors in student success, particularly in computing education, where high workloads, imposter syndrome, and performance pressure contribute to stress and burnout. This thesis aims to explore the challenges related to mental health in computing education and identify strategies to support student wellbeing.
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.
Main investigators for MIC-related projects are Frank Lindseth and Gabriel Kiss
It's also possible to focus more on model dev. for MIC, e.g. Vision Transformers (ViT), that can be tested on various organs and modalities.
Challenges related to MIC at key conferences like MICCAI and CVPR etc. can also be converted into interesting project proposals:
It's desirable to use MONAI (and MONAI Label) for all developments.
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
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.
Tasks can also be in connection to develop the modeling environment, and how to use AI to support modelling
A moving competitor is easy to segment when the camera is stationary. Once isolated, we can determine their pose, action and other characteristics. If a camera is also moving, it alone produces its own motion field that can be used to determine camera pose. Differentiating static from moving elements as well as camera pose is the problem of interest. We have had success in isolating and analyzing athletes playing ice hockey and baseball, determining their pose, action and other interactions. The goal is to extend this to sports where the camera also moves such as track, running, skiing, to name a few is the objective.
This project is under a collaboration with University of Waterloo, Canada and provides the possibility to take the pre-project at NTNU and the masters thesis at UofWaterloo. Taking the whole project at NTNU is also a possibility. Waterloo has data for ice hockey, baseball and other sports. Collaborations with other sports in Norway/Canada are also possible to obtain further data.
Det er i dag mange kunstig intelligenssystemer i bedrifter, alt fra enkle algoritmer, til kompleks bruk av språkmodeller. For å kunne utnytte de på best mulig hvis er det mulig å benytte seg av et Multi-agent-systemer (MAS), der man legger til rette for samhandling og kommunikasjon mellom både tekniske agenter, og mennesker.
Gjennom action-design research vil man i denne oppgaven kunne utvikle og undersøke effekten av et slikt system i praksis.
Du vil gjøre innledende litteraturstudier om emnet og designe en design science studie med kravinnhenting, utvikling og evaluering gjennom 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 av SINTEF Digital som jobber med denne problematikken.
Ta kontakt med Marius før du velger denne oppgaven
Architectural design of floor plans is a time consuming and labor-intensive task. Computer-aided architectural design can ease this work though automatically generated floor plans for office buildings can advance the research field of computer-aided architectural design.
The generation of such floor plans need to be feasible wrt architectural constraints.
Example previous project:
VI har en avtale med Arealize (https://www.arealize.ai/) i Trondheim om en enkel samarbeid på prosjektet. Arealize har arkitekts som er villig til å bidra med råd og noe testing/interaksjon med en simulaltor, ved behov. I tillegg har Arealize noe database for innredning som kan og slåes sammen med areal planlegging eller kan utvikles som en adskilt prosjekt om innredning av en floor plan.
En industri kontrakt ønskes ikke av Arealize. Da er studentene fri til å utvikle prosjekt i den retningen som ønskes men med mulighet for noe profesjonal råd fra Arealize, ved behov.
https://www.arealize.ai/
Studentene er fritt til å velge sin egen vri på tema inkluderte å bruke et annet biologisk inspirerte algoritme.
Merk at prosjekt er og knyttet til prosjekt tema innen Decision making for Multiobjective optimisation pga teknikk og behov.
Personality is a set of traits and unique characteristics that give both consistency and individuality to a person's behavior. As personality is accepted as an indicator of job performance, recruiters aim to retrieve these behavior traits in the screening process. One issue is that using personality questionnaires is less favored by applicants and negatively affects the pace of the recruitment process. Many recent studies started exploring asynchronous video interviews (AVI) and social platforms to predict one's personality. This study aims to explore and develop machine learning algorithms (preferably multimodal DNN) for analyzing recording interviews and accompanying resources (social media/online profile presence such as LinkedIn) in predicting one's personality on Big Five personality traits.
The objectives of this project is to develop a plant health monitoring system with IoT-enabled in-field soil sensor data and UAV images (RGB, Multispectral) to accurately detect, classify, and predict plant health.
The existing plant health monitoring systems are either rely on in-field below-ground sensor data (soil moisture level, nutrients level, pH level etc) or on visual assessment (images) however, plant health depends on both soil properties below the ground and visual symptoms develop above the ground. For instance, nutrient deficiencies or water stress in the soil often result in subtle visual symptoms on plant canopies that may be overlooked when evaluated separately.
Furthermore, the correct evaluation of plant health becomes complicated due to integration limitations of data sources whose characteristics differ in terms of spatial and temporal elements and noise charactristics. There is urgent need for a comprehensive multimodel plant health monitoring system that intregrates soil sensor data with UAV images to capture dynamic relationship over time to enable proactive and data-driven decision making in precision agriculture. This integrated approach will not only improve the accuracy of plant health assessments but also provide actionable insights to optimize resource utilization, enhance early intervention strategies, and promote sustainable agricultural practices.
Data Requirements
For this project we need Temporal infield soil sensor data (including labels) and UVA images. The soil sensor data is mostly extracted with RS485 Modbus Soil Sensor which includes soil moisture, pH level, salt level, nitrogen phosphorus, potassium level, and temperature. The UAV plant images, with labels. The publicly available datasets for plant disease detection include PlantVillage, NewPlant, PlantDoc. These datasets does not contains below-ground sensor data against each image.
The candidate will work in collaboration with Nibio for data collection if required.
Knowledge required: Image processing, Deep learning, AI
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
The project will mainly be based on simulations in our open-source flatspin simulator for large-scale ASI systems.
Music presents a unique way to view a dynamical system. If we can express a system's behaviour through music, then we have a new way to perceive and understand the system, as well as a novel method to create new music.
Artificial Spin Ice (ASI) are dynamical systems with a rich variety of emergent behavior. Recent work in the COMET group has demonstrated how the dynamics of ASI systems can be used to create music. First, an ASI is simulated and perturbed in some manner. Then the results are fed through a custom mapping to produce music. We also used evolutionary methods to search for an ASI geometry that produced the “best” music.
The aim of this project is to extending this work. This could be improvements to the mapping, allowing it to produce better sounding music, or music that more closely expresses the behaviour of the ASI. Alternatively, one could test out and develop different fitness functions, i.e., means of evaluating how good the produced music is, and the use this for an evolutionary search to find an ASI geometry with even better musical qualities.
Overview / Motivation
Neural Radiance Fields (NeRFs) have revolutionized scene representation by enabling high-fidelity 3D reconstruction from sparse views. Originally developed for static scenes, recent research extends NeRFs to dynamic and unstructured environments. For intelligent mobile robots operating in real-world missions—such as inspection, exploration, or multi-agent coordination—robust, real-time scene understanding is critical. This thesis explores how NeRFs can be adapted for environmental representation in dynamic settings using rich sensory inputs. Leveraging stereo vision, LiDAR, and IMU data, the goal is to build a scene representation that enables both localization and interaction in environments with other moving vehicles.
Thesis Proposal: NeRF-Based Dynamic Environment Representation with Multi-Modal Inputs
Problem Statement
In complex, unstructured environments—such as industrial sites, forests, or urban landscapes—autonomous mobile vehicles must build and update internal models of their surroundings for navigation, planning, and interaction. Classical mapping techniques (e.g., occupancy grids or SLAM) offer limited semantic richness and flexibility. NeRFs offer a compelling alternative by enabling implicit volumetric scene representation. However, challenges arise when dealing with real-time, dynamic environments and when integrating multi-modal sensor inputs.
Project Goal
To develop and evaluate a multi-modal NeRF-based system that fuses stereo camera, LiDAR, and IMU data to construct a real-time environmental representation suitable for autonomous agents in dynamic scenarios.
Approach
Expected Outcomes
Co-supervisors:
Computed tomography (CT) is an imaging technique used to reconstruct an object’s internal structure in a non-destructive way. In medicine, CT, alongside MRI, PET, and ultrasound, is used to create 3D models of a patient that can be inspected for injuries or other changes. CT is also used in earth science and material science to better understand the internal structure of the Earth and different materials. The goal of this project is to investigate improved methods for 3D and 4D reconstruction of CT-scanned objects by means of machine learning. A recent machine learning method known as implicit neural representation will be among the methods studied. We have found in already-performed experiments that implicit neural representation produces state-of-the-art results, surpassing the reconstruction quality of other methods both visually and quantitatively.
The project will build on existing research performed, including an award-winning MS Thesis “NeCT: Neural Computed Tomography for Sparse-View and 4D CT” (see below). The research project can go in different directions based on the interests of the project team, including new students, and results achieved at the end of the Spring 2025 semester.
Master's students Henrik Friis (Department of Physics at NTNU) and Håkon Nese (Department of Computer Science and Informatics at NTNU) won second place in the "NAIL Best Master Thesis Award 2024," organized by the Norwegian Open AI Lab (NAIL), for their thesis "NeCT: Neural Computed Tomography for Sparse-View and 4D CT". Their thesis explores how neural networks can be usedto significantly improve temporal resolution in computed tomography. By using the new program NeCT, dynamic 4D (3D and time) datasets can now be reconstructed with a temporal resolution on the scale of seconds instead of hours.
A team consisting of two MS students with partially overlapping, partially different strengts works well. One student may have a strength more towards physics and engineering, while the other student may have a strength more towards artificial intelligence and machine learning.
There is increasing awareness in society and in the scientific field regarding the downsides of children’s sustained engagement with screen-based systems.
Multisensory capabilities such as recognizing users’ body motions and gestures (e.g., using depth sensors, accelerometers, and gyroscopes), positioning recognition (e.g., using traditional radio-frequency short-to-long distance identification such as NFC or RFID), and speech recognition are now allowing children to interact in a more natural and multimodal manner. Generative AI (GenAI) creates new content and allows us to deliver tailored feedback and recommendations to children (e.g., through conversational interactions). Multisensory GenAI will have to function alongside other activities.
The aim of this project is to develop a paradigm of non-visual mixed reality
We wish to explore interactivity, where users engage with their full senses in the physical and social environment surrounding them while also interacting with virtual interactive elements overlayed to the physical space that does not rely on visuals. The intended benefits of such interaction are:
Through interaction design experiments conducted in various settings, we shall seek an optimal combination of haptics, audio, etc, to augment/better support children. Through elicitation studies, we shall seek to identify common requirements and adaptive solutions that enable natural and intuitive interaction.
As an example of the kind of interactivity we aim for, consider children playing a computer game where virtual game objects are spread in the physical space. These objects are to be experienced through ‘magic-wand’ like handheld devices and wearables that provide haptic, sound, and light feedback in response to movement and physical actions of players. Similar interactions can support learning and communication applications.
Ultimate Visual Computing and AI project.
Several sub-project, some examples:
We want to build a close to real-time photorealistic digital-twin of the area around the Gløs campus on top of Omniverse based on USD (or Unity / UE), as a start - can easely be scaled. The work has already started, useing existing data like terrain models, orthophotos, buildings, geo-located point-clouds and images for context, then update the model with real-time data to match the physical twin. Many of the tasks needed can be automated, for example 3D content creation from 2D images.
It should be possible to import the generated environment / model into simulators like DRIVE Sim, Isaac Sim and Carla and learn agents to drive itself (modular and end-to-end approaches like imitation and reinforment learning).
At the end we want to fine-tune the AI-models created in simulation with real data and apply the methods to our new AV platform for research on autonomous driving (AD).
Sounds interesting? Want to contribute? Let's discuss a good project fitting to your background.
The goal of this thesis is to produce knowledge about the state of the practice in Norway about development of AI intensive systems.
The paradigm of research will be empirical software engineering. The student(s) will analyze the material produced by the groups in TDT4290 and study how the different companies relate to AI and which are the trends. The students will run literature review (in the Autumn), interviews of customers and students, report analysis, and analysis of software.
Patón-Romero, J. David, Ricardo Vinuesa, Letizia Jaccheri, and Maria Teresa Baldassarre. "State of gender equality in and by artificial intelligence." IADIS International Journal on Computer Science and Information Systems 17, no. 2 (2022): 31-48.
Large language models (LLMs) will be gradually implemented on edge systems, like mobiles, self-driving cars, etc., as a powerful AI assistant. Normally, LLMs are trained to provide broad knowledge to different users. However, as learned from lectures on convolutional neural networks (CNNs), personalizing CNNs (removing some classes never used) can reduce the model complexity and latency while in some cases improving accuracy. The goal of this project is to evaluate an on-device personalization method for LLMs on edge systems. This means that the entire personalization procedure will be conducted on edge systems without the assistance of remote servers. Thus, the proposed method should be low overhead and hardware-aware so that the limited computational power of edge systems can be fully utilized to achieve this goal.
Denne oppgaven tilbys i samarbeid med MIA Health
I dag er MIA stort sett aktivitet (i form av puls og PAI) knyttet til forebygging, ink. muligheten for å koble seg til ulike wearables som smart-klokker og helse-monitorer, og det som trengs av frontend/backend og user-adm for å håndtere dette.
Kunne du tenke deg å være en del av et dedikert team som jobber i retning av en tvilling som passer på deg og hjelper deg gjennom livet? Noen forslag til problemstillinger som det går an å se på:
Forslagene over er akkurat det, dvs. forslag. Og kompetanse fra de fleste av IDI sine studieretninger vil være av interesse i denne konteksten. I et konkret prosjekt samarbeid etterstreber vi alltid at studenten er en del av diskusjonen og en del av forprosjektet vil bestå i å sammen spikre en master-oppgave som studenten virkelig er motivert til å jobbe med. Hvis ønskelig kan vi også avtale et møte med MIA før deadline for prosjekt ønsker går ut.
Progresso (previously ProTus - https://protus.idi.ntnu.no/, username: testUSN@usn.no, password: test) is an evolving online learning platform designed to provide learners with personalized courses across multiple domains. Over the years, different versions of the system have been developed, gradually expanding its capabilities to include new content, analytics, and personalization features.
The current iteration of Progresso offers courses in Java, integrating interactive third-party materials while tracking learner interactions and providing learning analytics. However, as learning methodologies evolve, there is a need to incorporate Self-Regulated Learning (SRL) principles to improve student engagement and autonomy.
This thesis will focus on the further development of the platform by implementing a Self-Regulated Learning module that supports students in setting learning goals, monitoring progress, and adapting their strategies based on feedback. Additionally, a user study will be conducted to evaluate the effectiveness of this approach in enhancing learning outcomes.
The project consists of the following phases:
The ideal candidates will have a background in software design, solid programming skills and an interest in hands-on development and experimentation.
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.
The content can be adjusted based on the interests and background of a given student. But some interesting research questions to look into could be:
Some links: Lamina.ai with blog, LLMsPracticalGuide, and natural conversation.
The project can be linked to MIA Health if interesting for you.
A very recent paper supporting our initial thoughts and ideas behind the overall concept can be found here:
"Patients prefer ChatGPT over Physicians 79% of the time! A recent experiment used ChatGPT to answer medical queries on an online forum (Reddit). The responses were later graded by medical professionals and compared to responses from actual human physicians. As you can see in the charts below, the "robots" produced better-quality answers, and their answers also displayed more empathy."
Furthermore, the Med-PaLM 2 model just released scored 85.4% on the US Medical License Exam (USMLE) questions, i.e. expert knowledge (which is a 18% leap over previous state of the art results from Med-PaLM some six months ago, the first AI model that passed USMLE).
Interested? Get in contact so that we can discuss and design a project that would be really motivating for you to work on.
Life on earth displays an explosion of creativity and diversity. One single run of evolution has managed to create both photosynthesis, flight and human intelligence and is still presenting us with new solutions to the problem of survival and reproduction on earth. This capability for never-ending creation of novel organisms is what the field of open-ended evolution is trying to replicate.
One such open-ended evolutionary algorithm is the Paired Open-Ended Trailblazer (POET) algorithm. This is a coevolutionary algorithm seeking to endlessly generate problems of increasing difficulty and their increasingly complex solutions through the enforcement of a minimal criterion and goal-switching. POET has been applied to robot locomotion control and general game-playing environments.
A recent masters project has focussed on a new hyrbid algorithm involving POET and Neuroevolution of Augmented technologies (NEAT). An ablation study was performed to determine whether the algorithm’s minimal criterion and transfer mechanisms are in fact necessary, in thte light of a more challenging control problem with maze-like characterisitics with the introduction of adversaries.
This project is suitable for extensions, as is ie with such a hybrid algorithm as the basis, or to choose a different algorithmic approach to investigate open-ended evolution.
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.
With the increasing use of video surveillance of public spaces, there is a growing need to take care of the privacy and integrity of those who are filmed by the surveillance cameras (Asghar et al. 2024). Full encryption of the video removes the ability to quickly find out what is happening on the video. The identity of persons who are not relevant to any investigation will then also be known by decryption. One solution to this problem is to partially encrypt the video.
In this proposal for a bachelor's project, we propose to encrypt parts of the video stream. There may be individual blocks in the video stream that contain faces or other things that can identify people. The main challenge in this project is to find a solution that does not require new players for the videos. The encrypted code must be inserted into the video stream in such a way that it is not corrupted at the same time that the hidden information cannot be extracted by unauthorized persons. To simplify the problem, one can choose to either encrypt, for example, license plates or faces.
The main objectives of this assignment are as follows
Particle Swarm Optimisation, where each particle may be thought of as an individual in a society of particles, provides a technique to model influences and effects in society that may prove beneficial in the battle against Climate change. This project is focussed on developing a PSO modelling environment that enables the study how an individual may become more inclined to increase their individual contribution/sacrifice for the benefit of future populations.
contact: Pauline Haddow
The human body constitutes a complex system wherein a large collection of parts interact. Consequently, medical interventions have varying effects on patients. Personalized Health is the field of research striving to predict strategies for medical interventions optimized for individual patients. In the scope of Personalized Health, recommender systems play a central role. They take a vast set of option and automatically narrow them down based on a user profile. The master project will pick a specific problem and implement a recommender system to personalize interventions. Candidate problems include, but are not limited to, nutrition, exercising, medication, and therapy.
The candidate should have attended courses about Artificial Intelligence, Machine Learning, and Recommender Systems. Ideally, the candidate has a vested interest in the topic. The candidate should have sufficient programming skills to process the underlying data, implement the algorithm, and develop a demonstrator. Experience with Recommender Systems and medical data is a plus.
Many patients need personalized training videos to perform rehabilitation at home. The current training videos from therapists are standardized, and do not fit individual needs.
This project aims to use sensor data and Generative AI technologies to generate personalized training videos based on the exercises the patients practice during the clinical investigation.
The research questions are:
The project will be co-supervised by Prof. Frank Lindseth and Associated Prof. Gabriel Hanssen Kiss.
Several models can be used to find out how users’ social media networks, behaviour and language are related to their ethical practices and personalities, Such models include Schwartz’ values and ethics model and Goldberg's Big 5 model that defines personality traits such as openness, conscientiousness, extraversion, agreeableness and neuroticism. The thesis project would investigate applying such models to social media text and how the user personalties are reflected by the social networks that they participate in and develop.
This project is linked to an EU project that deals with climate change and its effect on biodiversity in Sea. The partner company in this project, Synplan (Oslo based start up, https://www.synplan.ai) will co-supervise this thesis.
In the EU, nearly half of the population lives less than 50 km from the sea. These coastal populations are continuously growing, increasing anthropogenic pressures on marine ecosystems. Predicting spatial and temporal biodiversity dynamics—an essential component of mitigation strategies against human and climate change-related impacts—has become increasingly urgent and vital.
This master’s thesis focuses on AI-based methods (CNNs, Transfer learning, Resnet, VGG...) for curating datasets necessary for such predictions. Specifically, it involves the use of images of marine species collected using both low-cost tools (e.g., Planktoscope, Lamprey-MultiBarcodeTools) and high-end instruments (e.g., FlowCam, Cytosense), as well as satellite data. These images need to be labeled—i.e., the species of phytoplankton present in the images must be identified. While scientists currently label some of the data manually, the process is time-consuming and labor-intensive. To scale up the dataset, there is a clear need for AI/ML-based image recognition methods. The project may also involve integrating data from multiple sources, such as frugal and high-end instruments, to enhance the robustness and applicability of the models.
The project fits to one or two students.
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:
Probabilistic models for generative AI like diffusion models (see e.g. these lecture notes) have gained quite some interest lately, with impressive results in particular in the image domain (now also doing high quality video from text), but also used for a lot of other things including reinforcement learning and generation of both time series data and graph data. Diffusion models have also inspired more fundamental research into alternative formulations of generative AI, including Bayesian flow networks and flow matching models.
In this project you will work with models for generative AI. The content of the project can be adapted to your interests: A purely theoretical task, a comparative study (e.g., a shootout between different versions of diffusion models, comparing to GANs or VAEs, etc) or a practical one related to a dataset of your interest are all possible.
If you are interested in this project, please make sure to look at this page.
The bioinformatics group works on developing and using computational models to predict how changes in gene regulation can control development and cause disease. Towards this end, we develop custom algorithms, statistical simulations, and machine learning-based solutions to analyze and interpret biological data; examples of previous MSc-theses include a genetic programming (GP) approach to predict microRNA target sites, a support vector machine (SVM) approach to identify microRNA genes, and an approach that combines GP and SVMs to identify related proteins.
We have several project and MSc-thesis opportunities for students interested in programming, machine learning, and string and search algorithms. If you are interested, please contact me to discuss specific projects.
I Kredittbanken har vi saksbehandlere som bruker mye tid på å lese og vurdere dokumentasjon sendt inn av kunder, i forbindelse med søknad om kredittkort, refinansiering eller forbrukslån. Dette arbeidet består ofte av å hente ut informasjon, kontrollere innholdet og gjøre vurderinger basert på det som står i dokumentene. Dokumentene er typisk dokumentasjon på inntekt, leiekontrakt, betalingsinformasjon i forbindelse med oppgjør av lån. Dette arbeidet er tidkrevende og sårbart for menneskelige feil, og vi tror det finnes potensiale for delvis eller full automatisering av prosessen.
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.
The overall objective of this project is to train deep learning models to detect and identify students’ submissions to their own work vs copying and the use of AI-assisted tools by profiling students’ submissions. The candidate is required to train attention-based deep learning models, or alike, to learn and identify writing patterns that differ from one’s own work. The candidate is expected to conduct a literature review on the topic to identify which models can best classify/identify individuals writing styles based on prior data, what features/patterns are important to track, and how to best use them for handling submissions. The candidate is expected to design use cases to train and test the system. Experiments can be designed to collect summary/reflection notes on the topic by a group of students to collect datasets for the thesis work.
KANs & GP m/ Signe på SINTEF
Denne masteroppgaven har som mål å utvikle en modulær og utvidbar Python-basert analysepipeline for kvalitativ data, med spesielt fokus på intervju-transkripsjoner. Studenten skal først kartlegge state-of-the-art metoder innen kvalitativ dataanalyse med digital støtte – inkludert naturlig språkprosessering (NLP), tematisk analyse, nettverksanalyse og visualisering. Deretter skal studenten utvikle og dokumentere en prototype-pipeline hvor man kan mate inn transkripsjoner og få ut relevante innsikter, som f.eks.:
Oppgaven følger action-design-research prinsipper og vil innebære og undersøke effekten av et slikt system i praksis med forskere som bruker denne type data.
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 SINTEF.
Ta kontakt med Marius før du velger denne oppgaven.
While many students use generative AI tools such as Microsoft Copilot, we have little knowledge of how these tools are used, for what, and how they affect student learning. In this project, you study the use of Microsoft Copilot by computing students at NTNU using qualitative research methods to gain a rich understanding of the phenomenon. You will do initial literature studies on the topic and design a case study with data-collection methods like observations, interviews, and archival data. You will analyze the collected data qualitatively to explaining the practices of computing students using Copilot.
This master thesis aims to test and evaluate various off-the-shelf state-of-the-art large language models (LLMs) for Retrieval-Augmented Generation (RAG) tasks on domain-specific education content for course preparation. The student is expected to test various models on benchmark datasets and design the experiment for a real-case scenario. The material and content will be provided as a use case. The candidate is also expected to validate the responses using objective and subjective evaluation criteria. The effectiveness of such a tool may also be studied in this master thesis by gathering students’ feedback and conducting surveys on experimental groups. Prior knowledge: Machine Learning, Generative AI, LLMsSkills: Python programming
To maximally exploit existing infrastructure while not risking power outages, Statnett wishes to develop models for detecting anomalous behaviour of their voltage transformers. Currently, a straightforward algorithm is applied, detecting drift based on differences between values per timestamp and a given threshold. Ideally, drift should be detected early and time to a given threshold should be predicted.
The candidate will have access to time series of voltage for three phases for more than 1000 transformers. The data are unlabelled, but deviations between the three phases indicate a need for maintenance (can be on short or long time scales), however, data is of varying quality and quality issues can mimic drift. Systems with many externalities can be considered dynamic and thus represent a challenge for data-driven AI. The candidate will explore ways to discover rare anomalies coinciding with the need for maintenance.
The project will be co-supervised by Signe Riemer-Sørensen (SINTEF), Abdul Kazeem Shamba (NTNU), and Barbara Barzycka (Statnett).
The gold standard for computer graphics is and always has been the simulation of real light dynamics through ray-tracing, but due to the high demands on compute it has seen little use in real time applications. In recent time ray traced real-time graphics has become a reality, with the last year even bringing ray-tracing capabilities to mobile devices. The advances that made this possible are several, including better process nodes for silicon, advances in neural network based denoising, novel temporal antialiasing techniques and improvements in Bounding Volume Hierarchy (BVH) construction. Ray tracing relies heavily on specialized data structures to make the intersection test between the ray and scene efficient using some kind of Acceleration Structure, with the standard approach being the use of a BVH. The BVH is a spatial data structure, and lends well to GPU warp-based execution because rays issued from nearby pixels (and scheduled on the same shader core) are likely to traverse the same part of the tree. This allows the tree nodes to be re-used for multiple threads of execution saving an order of magnitude in bandwidth. The difference between a well- and poorly constructed BVH can account for more than 50% of the ray traversal performance making quality a very sensitive topic. Another issue is that higher quality build algorithms naturally require more time, to the point where building the BVH takes too long to be feasible in a real-time environment. Due to this tradeoff between traversal performance and build time the field of acceleration structure construction is wide open and there are multiple heuristics that can be applied in attempt to get ahead in one way or another. The complexity of the problem is further increased by the fact that different hardware accelerators have different performance characteristics, meaning the same algorithm may not be the best everywhere. This all means that the construction of BVHes is not in general well understood, and there is ample room for innovation.
For the Fall project we would have you implement a standard Surface Area Heuristic (SAH) or Linear BVH (LBVH) algorithm and insert the resulting BVH in ARM’s hardware to evaluate performance. The initial implementation should be on CPU since it is much easier to program, but depending on how it goes you are encouraged to implement a build algorithm using Vulkan Compute as well. The goal is for you to have a solid base (and test pipeline) ready for your Masters thesis project, where you will have the opportunity to explore the various ways in which the standard algorithms can be improved upon. The specific direction here is up to the student. The algorithms that are implemented can be evaluated on their build time, their memory footprint and on their effect on the framerate of sample content. Several competing algorithms have implementations available on github and can be used for comparison.
Suitable for: 1 studentSupervisors: Theoharis Theoharis, NTNU, Torbjörn Nilsson, ARMRequirements: Computer Graphics courses (see below), Knowledge of C++, OpenGL, interest in learning VulkanCourses: TDT4195 (Visual Computing Fundamentals), TDT4230 (Graphics & Visualization), or equivalent.
Literature:1. https://jacco.ompf2.com/2022/04/13/how-to-build-a-bvh-part-1-basics/2. http://www-sop.inria.fr/members/Stefan.Popov/media/KDTConstructionGPU_TR10.pdf3. https://www.nvidia.in/docs/IO/77714/sbvh.pdf4. https://meistdan.github.io/publications/ploc/paper.pdf5. https://devblogs.nvidia.com/wp-content/uploads/2012/11/karras2012hpg_paper.pdf6. http://gamma.cs.unc.edu/SATO/SATO_files/sato_preprint.pdf7. https://vulkan-tutorial.com/8. https://research.nvidia.com/sites/default/files/publications/dnn_denoise_author.pdf9. http://behindthepixels.io/assets/files/TemporalAA.pdf
Car tires have different features relevant for maintenance. These features include the profile depth, size, and manufacturer. Today, organizations tasked with tire maintenance need to keep track of these meta-data manually. In other words, employees have to inspect the tire and notes down existing damages as well as the meta-data.
The project's objective is to automate the process with the help of computer vision technology. The candidate will get access to a large collection of images of tires from different angles and meta-data that has been manually collected. Subsequently, the candidate will build computer vision models to predict the correct meta-data, estimate the profile depth, and detect possible damages.
The project is supported by Andreas Rosmo from Trønderdekk AS.
A recent addition to the modeling of scenes is based on 3D Gaussian primitives. The associated rendering technique called Gaussian Splatting:
https://en.wikipedia.org/wiki/Gaussian_splatting
The idea behind this project is to create a 3D Gaussian representation of the inside of Nidaros Cathedral, based on a large set of photographs taken by Torbjørn Hallgren.
Following, novel views and walkthroughs will be generated.
Knowledge: Python, C/C++
Prof. Theoharis Theoharis, IDI, NTNU theotheo@ntnu.no
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
Approximate computing studies how to provide “good enough” results for a certain application. It is used in different context, for example resource constrained devices or when operating in degraded mode. We want to evaluate the impact of faults on different approximate computing techniques.
Approximate computing [1] 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 certain performance metric such as time-to-solution, energy-to-solution, etc. Many approximate computing techniques exist. In this project, we focus on compiler-level techniques.
Fault injection [2] is a verification technique that deliberately introduces faults into a system, to evaluate their effects. It is often used in the testing of critical systems, to create conditions to test redundancy and recovery mechanisms. Different techniques may be used. In this project we focus on injection at code level and during the compilation process.
While different approximate computing approaches may produce similar results in terms of quality and performance metrics, they may have very different behavior in terms of their ability to cope with faults (fault tolerance). The objective of this project is to apply fault injection to compare different approximate computing algorithms. This project will be based on existing approximate computing algorithms and existing benchmarks. The main task is to extend such benchmarks with a fault tolerance perspective, in which faults are injected to the existing code base.
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?
Evaluate different object detection and/or trajectory planing algorithm from the safety perspective. Involves experimenting with different ways to evaluate object detectors, and possibly defined new benchmarks. Builds on existing research and a previous Master’s theses at IDI.
Autonomous vehicles rely on object detection as a fundamental way of perceiving the environment. Modular pipelines for autonomous vehicles first acquire data from sensors, perform object detection, create a scene representation, and finally perform motion planning.
Autonomous vehicles are safety-critical systems [1], where “safety” is defined as the absence of catastrophic consequences on the users and the environment. However, machine learning components are typically evaluated with traditional metrics based on precision and recall.
Recent works in the literature have proposed metrics and algorithms for object detection that integrate the concept of safety (e.g., [2]). The objective of this project is to evaluate object detectors from the safety perspective. The task involves a literature review on safety metrics for trajectory planning on autonomous vehicles. This work builds on existing research and on a previous Master’s thesis that has focused on the object detection step [3].
The long-term research objective linked to this activity is to define a methodology that can assess the autonomous vehicles from a safety perspective.
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.
Football has become an enormous industry where players are bought and sold based on information about them. This information is not just collected during games, but during trainings, tests, leisure and even when sleeping. In the wrong hands, this information can be used for extortion, manipulate the market or influence betting odds.
In 2022, the player representative organisation FIFPRO launched the the Charter of Player Data Rights. This charter aims to protect the privacy of professional footballers and allow them to benefit from personal rights to manage and access information about their performance and health. However, there is a lack of available implementations for the protection, collection and use of player data. For instance, players do not appear to have sufficient means to control the collection of their data and they are also concerned with the portability of moving their data from one club to another. Also, such solutions should become available and affordable not only to top players, but amateurs and youths as well.
This project will investigate how technical solutions related to distributed storage and data governance can support player data rights, and at the same time support positive usage of this data (improve player development, prevent injuries, engage fans, economic compensation). The student(s) will do a literature study and develop threat models to potential solutions. In a master thesis continuation, a proof-of-concept should be developed and tried out with representative end users (e.g. football players).
Note: The specific details of the assignment can be tailored according to the student's interests and expertise. A previous master project on this topic has been conducted in the spring of 2025, and this assignment will build upon the results.
Human activity recognition refers to the task of automatically identifying and classifying different activities that a person is performing, such as walking, running, sitting, or standing, based on data collected from various sensors.
The Transformer model, originally introduced for natural language processing tasks, has shown remarkable performance in capturing long-range dependencies and patterns in sequential data. This project focuses on developing the Transformer architecture to process and understand sensor-generated sequences, which could include data from accelerometers, gyroscopes, and other wearable or environmental sensors.
DNV ønsker å bruke sensorer for å automatisere klassifisering, redusere manuelle inspeksjoner og sikre kontinuerlig overvåking.
En utfordring er datadeling og standardisering av forskjellige formater av data til forskjellig bruk. Noen skip har sensorer, men deler ikke data, mens andre ønsker standardisering, men mangler teknologi. VISTA Gateway skal samle og behandle sensordata, men krever skalering og støtte for flere dataformater. Maintenance Activity Data (MAD) følger ikke ISO-standarder, noe som gjør standardisering krevende. Oppgaven kan undersøke hvordan ulike standarder kan kombineres, og hvordan sanntids- og batchprosessering kan optimaliseres.
Oppgave kan ta for seg potensialet for virkningen en implementering av sensorbasert klassifisering vil ha, i form av tekniske behov for lagring og prosessering, organisatoriske behov med hensyn til intern og eksterne forhold og bærekraftspåvirkning en implementering av sensorene vil ha.
Social media posts often express sentiment (positive or negative emotions) towards a product, person, political party, etc. The project is aimed at the automatic classification of sentiment in texts on social media, tentatively addressing issues involving negation and/or figurative language, that is, language which intentionally conveys secondary or extended meanings (such as sarcasm, irony and metaphor). Figurative language creates a significant challenge for sentiment analysis systems, as direct approaches based on words and their lexical semantics often are inadequate in the face of indirect meanings. A subgoal of the project would then be to find a set of social media posts rich in figurative language, but the main goal would be to determine whether a user has expressed a positive or negative sentiment and possibly the degree to which this sentiment has been communicated.
Shape servoing is a form of robotic manipulation that involves altering the shape of a deformable object, such as soft plastics, fabrics, or muscle tissue. Manipulation of deformable objects is generally more challenging than rigid objects but successfully doing so may unlock transformative changes in several industries, including manufacturing, agriculture, and medicine.
In a previous work* (Herland and Misimi 2025), we developed a shape servoing framework based on Deep Reinforcement Learning (DRL) that utilized discrete actions to manipulate 3D objects constrained to a 2D surface, thereby limiting deformation to two dimensions. This thesis aims to extend shape servoing to full 3D deformation, where objects can be manipulated in a more complex manner. Specifically, the objective is to train a DRL agent to deform objects that are fixed on one side and grasped from another, enabling them to achieve predefined 3D shapes. For example, an elongated deformable object could be manipulated to form different half-parabola-like structures or other desired shapes.
Objective
The system will employ an eye-to-hand configuration, where a camera observes the scene from the side and provides feedback to the robot about the desired deformation. The DRL agent must learn a continuous control policy to achieve the target shape while adapting to the material properties and constraints of the object. Key challenges include defining an effective reward function for 3D shape matching, ensuring stable control of elastoplastic deformations to achieve the goal, and handling variations in object stiffness and elasticity. The focus will be on the use of one to two particular elongated objects since the aim is the variety of possible 3D shapes achieved with a particular object(s) rather than in the diversity of objects it is capable of shape servoing.
This thesis will build upon the previous work on DRL-based shape manipulation, but will extend it by incorporating 3D deformation learning, continuous action spaces, and more complex object interactions, with potential applications in robotic soft object manipulation and deformable object modeling. The student will benefit from an existing framework and robotic setup at SINTEF Ocean. The MSc thesis entails signing an agreement with SINTEF Ocean.
*-the paper will be available at the start of the semester in September
Main Supervisor: Kerstin Bach (NTNU)
Co-supervisors: Ekrem Misimi, Sverre Herland (SINTEF Ocean)
Languages change rapidly over time and language users adapt to different situations and setting. Studying how language and communicative processes evolves is a highly multi-disciplinary task involving machine and language learning as well as biological and cultural evolution. The aim of this project will be to use methods such as evolutionary algorithms or reinforcement learning to investigate the main dynamics in language evolution.
When trying to understand the origins of languages, we can compensate for the lack of empirical evidence by utilising evolutionary computational methods to create simulations of how language may have evolved over time, e.g., by creating "language games" to simulate communication between agents in a social setting. In general, simulations on language evolution tend to have relatively small and fixed population sizes, something this study could aim to change.
A lot of work is being done on improving 3D shape matching algorithms for applications such as self-driving cars. A benchmark is needed that evaluates their capabilities in an objective manner that is representative for how these methods are going to be used in the real world.
A large variety of devices and methods exist for capturing such 3D data. For example, there is LiDAR, photogrammetry, infrared dot projectors, and structured light. Each of these has various tradeoffs between speed, the quality of the produced mesh, and the cost of the device.
Unfortunately, producing real 3D data with various capture devices for the aforementioned benchmark is costly and ineffective. First of all, the time required to capture each individual 3D model takes far too long to be reasonable for the number of objects you'd want to have for a benchmark that is reasonably reliable. Second, you generally want to test a matching algorithm under many different conditions such as noise, occlusion (when you can only see a portion of the surface), and clutter (other crap being present in the vicinity of the thing you're looking for). The requirement to store each of those variants on disk is the third issue, as this would require many terabytes of data.
A more scalable and economical approach is to use an existing dataset of 3D meshes, which you apply modifications to that simulate the same artefacts as those you get from various 3D scanning methods. As an added bonus, this allows the effect of each artefact to be tested separately in isolation. This would not be possible with real scans, as things like noise are always present.
This project aims to create a set of algorithms that alter existing meshes to have artefacts of various common 3D scanning techniques.
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.
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.
Smidige metoder oppstod i team som jobbet med systemutvikling, men har blitt tatt i bruk av større deler av organisasjoner og ført til endringer i ledelse av prosjekter og større programmer. Smidige metoder legger vekt på beslutningstaking på lavt nivå i selvstyrte team, tett kontakt med kunde og fleksibilitet i arbeidsprosesser for å kunne håndtere endringer underveis i utviklingsprosjekt. Når organisasjoner ønsker å ta i bruk denne type metoder i større deler av organisasjonen støter de ofte på utfordringer med eksisterende hierarki, at det kan være vanskelig å bli enige om arbeidsprosesser med miljø med andre oppgaver enn programvareutvikling og generalt at det er motstand mot organisatoriske endringer. Denne oppgaven vil bestå i å først gjøre et litteraturstudie på temaet “smidig transformasjon” innen fagområdene systemutvikling og informasjonssystemer, og koble med relevante andre felt som for eksempel endringsledelse. En videreføring i masteroppgave vil kunne innebære et empirisk studie av en eller flere organisasjoner som gjennomfører en “smidig transformasjon”.
For en innføring i forskningsfeltet, se forskningsagenda publisert på workshop på den internasjonale konferansen om smidig metodikk: https://link.springer.com/chapter/10.1007%2F978-3-030-30126-2_1
En masteroppgave kan delvis basere analyse på funn fra kompetanseprosjektet Agile 2.0, hvor NTNU var en partner sammen med DNV GL, Equinor, Kantega, Kongsberg Defence & Aerospace, Sopra Steria, Sticos. Prosjektet ledes av SINTEF Digital og er støttet av Norges forskningsråd.
Software architecture is a critical aspect of designing and developing software systems. Modeling and documenting the software architecture is a fundamental task in software engineering, and established modeling languages (e.g., UML) have been used for this purpose. This project investigates languages and patterns for modeling software architectures that include AI components.
Defining and documenting the software architecture [1] of a system is one of the most important tasks in developing a software system. Different selection and organization of components and patterns [2] have a large impact of non-functional attributes of a system, such as reliability, security, performance, etc. Over time, multiple methods have been defined to guide software architects in the definition of the most appropriate architecture for their system.
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 [4]. Despite having some differences, all these languages have been conceived for traditional software and systems. The objective of this project is to investigate the limitation of those languages for what concerns the modeling of software architectures that include AI components. For example, concepts like AI models, prompt patterns, AI agents, fine tuning, etc., are not explicitly captured by traditional ADL.
The long term objective of this project is to define modeling languages and patterns to specify software architectures that include AI components.
We want to start a research centre (Norwegian Centre of Excellence) that investigates the relation between Software Engineering, Artificial Intelligence, and Intersectionality.
The aim of this thesis is to explore the current practices in Norway in this field and to contribute to the theoretical and practical basis of this exciting and relevant area of research.
Software is used by everyone, but software does not, at present, provide equal rights and opportunities for all. Biases, which are prejudices for or against one person or group, influence both software development and the composition of the software engineering workforce. Software programming is increasingly supported by Artificial Intelligence, and there is a risk that existing biases will be perpetuated and reinforced if they are not properly researched and understood.
Examine how companies incorporate Software Engineering, Artificial Intelligence, and Intersectionality into recruitment and software development processes.
Investigate how software development teams are aware of and relate to Artificial Intelligence and Intersectionality.
Cryptic crosswords are puzzles that rely not only on general kwowledge, but also on the solver’s ability to manipulate language on different levels according to Sadallah et al., 2024 https://arxiv.org/pdf/2403. 12094. This is a hard problem, which requires much reasoning on the part of humans. The authors report an accuracy of 8.9% for the best Large Lan- guage Model (LLM), wheras human performance is 99%.
As a consequence, this application area is a great opportunity to explore the potential of LLMs in solving difficult reasoning tasks.
In this project, candidates will be refining the methods employed above, by benchmarking the LLMs produced at NorwAI agains the state-of-the-art models used in the above paper. Datasets for English also exist, but producing datasets for Norwegian is a second goal of the project. This can be done synthetically, and also in cooperation with rights’ holders.
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
Ruteproblemet, kjent som "Vehicle routing problem (VRP)" , har vært forsket på i over seksti år på grunn av dets teoretiske kompleksitet og betydning i mange bruksområder. Mange applikasjoner i det virkelige liv har et økt behov for å levere til tusenvis av kunder, mens de samtidig tar flere begrensninger i betraktning. I tillegg er applikasjoner i det virkelige liv ofte avhengig av å motta løsningene fra problemet raskt, noe som betyr at problemet må løses på kort tid. Likevel har få studier hittil fokusert på å løse instanser av stor skala med flere begrensninger, innenfor kjent tidsgrenser.
Et annet utfordring er dynamisk ruting utfordringer hvor mlijø er ikek ststisk, behov forandres raskt og ny løsninger trengs raskt, kanskje i sanntid.
Studentene kan velge teknikk og valgt retning innen temaet.
One of the core competence areas of professional involved in the development of IT is the identification and representation of requirements. For modern IT-solutions sustainability can be considered a key concern, thus we have to look upon how to achieve sustainability by design, and not as an afterthought. This project will investigate areas related to
• How to express sustainability requirements to software and information systems• How to elicit sustainability requirements to software and information systems• How to balance sustainability requirements and more traditional form of requirements• How to deal with sustainability requirement in modern agile development methods• How to evaluate adherence and achievement to sustainability requirements on a short and long time period
Typically the task will start with a structured literature review in the autumn, with a follow up in the spring evaluating the use of existing or new techniques for handling sustainability requirements in agile software development. The task will be related to NTNU Center for Sustainable ICT (CESICT), and the national network goforIT (Grønn omstilling for IT-bransjen)
Cloud computing is increasingly popular due to its potential benefits in scalability, cost savings, and flexibility. However, it also poses the risk of vendor lock-in, where an organisation becomes dependent on a specific cloud provider's proprietary technology. This dependence can again limit flexibility, increase costs, and hinder migration of software defined infrastructure to more sustainable cloud solutions [1]. To enhance the robustness of their systems, businesses are also looking for multi-cloud supported solutions.
To address these challenges and ensure the sustainability of cloud infrastructure, this master's project will investigate methods and strategies to mitigate the risk of vendor lock-in. The project will review the current state-of-the-art in sustainable cloud migration and explore practical solutions to reduce vendor dependency.
[1] J. Opara-Martins, R. Sahandi and F. Tian, "Critical review of vendor lock-in and its impact on adoption of cloud computing". i-Society 2014. doi: 10.1109/i-Society.2014.7009018.
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.
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.
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
The TechLARP project aims to close the gender gap in technology studies and STEAM education in order to encourage young girls to pursue STEM and, particularly, computer science studies.
TechLARP project aims to develop an innovative, arts-based Tech Education programme that draws on LARP (Live Action Role Playing), creative coding, and wearable technologies to build the capacities of student teachers to design and implement effective tech education interventions that increase young girls’ engagement, motivation, and confidence to engage with technology and broader STEM studies more effectively.
The goal of this thesis is to develop and validation of creative Tech inspirational resources such as videos, story cards, and presentations showcasing possibilities in computer science.
The student(s) will have the possibility to influence the integration of wearable technologies, AI technologies, Virtual Reality, LARP costumes, and community building activities with mentors and role models.
Høiseth, M., & Jaccheri, L. (2011, October). Art and technology for young creators. In International Conference on Entertainment Computing (pp. 210-221). Berlin, Heidelberg: Springer Berlin Heidelberg.
Papavlasopoulou, Sofia, Michail N. Giannakos, and Letizia Jaccheri. "Empirical studies on the Maker Movement, a promising approach to learning: A literature review." Entertainment Computing 18 (2017): 57-78.
https://www.techlarp.eu/about/
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.
There is considerable confusion about what AI brings of value for whom. This confusion is partly related to what we define as AI, but also because of the rapidly changing nature of AI technologies. In order to create a sustainable process of taking advantage of AI we need to have a framework for how we evaluate AI, and who benefits from AI and how.
In this task you will do a systematic literature review regarding the value of AI for society, organizations, and individuals. Based on the results of such an study you will engage in an empirical field study to generate new knowledge about the value of AI. The focus and the field for your study will be decided together with you. This task is within the discipline of information systems, and will rely heavily on literature from information systems, computer-supported cooperative work, and human-computer interaction. You are expected to create new empirical knowledge based on earlier research and own data generation.
One third of the food produced in the world today becomes food loss and waste. Grocery stores have implemented measures achieving reduction; however, there is still room for improvement both for reduction of food loss and for better utilization of unavoidable food loss (using them as animal feed instead of sending to biogas production). Increasing knowledge about food loss trends and properties can help to achieve these goals.Currently the sales data are used to assist in placing new orders to match the demand and avoid food loss. Furthermore, food loss data is reported as a percentage of turnover without considering the volume and quality even though these are available. Therefore, time series analysis and forecasting of the available food sales and loss data (weekly, monthly or holiday trends) can provide insights that can enable development of new reduction measures and increased information about properties of the food loss can contribute to increased value creation.The project is in collaboration with Cansu Birgen from SINTEF Ocean as part of the project called “Mapping food waste from Norwegian grocery stores”.
Traditionally, software development and operations have been handled by separate teams. In such an organisation, a central objective of the development team is to introduce new application functionality, something that conflicts with the operation team's goal of ensuring stable and reliable services. The conflict in terms of the functions and values that the two teams provide often lead to the implantation of rigid processes that delay development. Hence, recent approaches aim to break down these barriers, forming multidisciplinary teams (DevOps) that are responsible for the entire lifecycle of an application, including operations and maintenance. Organising people from both development and operations in such teams, while also automating parts of the deployment process, has showed to both enable rapid deployments of new functionality and to increase software quality [1].
However, many organisations still struggle to adopt DevOps practices [2] and transforming traditional software and IT operations into DevOps teams is not straightforward. This project will investigate how organisations have managed the transition to DevOps, and how the different functions traditionally provided by an IT operations team can be re-organised in a DevOps development model.
[1] Forsgren, N. et al. Accelerate: The Science of Lean Software and DevOps: Building and Scaling High Performing Technology Organizations. 2018. [2] Krey, M. et al. DevOps adoption: challenges & barriers. HICSS2022. https://digitalcollection.zhaw.ch/bitstream/11475/24225/3/2022_Krey-etal_DevOps-Adoption_HICCS.pdf
Problem description TrønderEnergi has several forecasting systems utilizing machine learning methods in commercial operation. The systems are used and the forecasting is monitored by domain experts in the operating center at Berkåk. The domain experts do not have any training and knowledge in the machine learning methods, but they do have to make decision based on their output. At times the experts need to understand why a forecast has the value it has. The literature has to be reviewed, prototypes should be developed and evaluated by domain experts.
Data There are several data sets that can be used for this including consumption, grid loss, weather forecasts, and power production data. Many of the data sets have more than five years of data.
The focus will be on data from deployed systems, so that they can be improved.
Challenges Provide explanations to domain experts, so that they understand the forecasts.
This project aims to address the challenge of Uncertainty Quantification (UQ) for Language Models [1], recognizing its crucial role in ensuring confidence and reliability in the generated text.
The increasing adoption of Language Models in safety-critical applications highlights the pressing need to assess and quantify the uncertainties inherent in their outputs. While Language Models have demonstrated remarkable capabilities in natural language generation, their outputs often lack transparency, raising concerns about their reliability in critical scenarios. Uncertainties in Language Models can stem from various sources, including data limitations, model architecture, and the inherent stochastic nature of language generation.
Ensuring the reliability of Language Models in safety-critical systems requires effective methods for uncertainty quantification. However, achieving robust UQ for Language Models presents significant challenges. Existing approaches often lack scalability, struggle to capture nuanced uncertainties, or require extensive computational resources. Additionally, the interdisciplinary nature of the problem necessitates a thorough understanding of both language modeling techniques and uncertainty quantification methodologies.
The student is expected to review relevant literature on UQ for Language Models to gain insights into the strengths and limitations of current UQ techniques and identify promising avenues for further research. The student will implement and test selected Language Models (e.g., deep learning [2], Tsetlin machines [3], probabilistic models [4]) and UQ methods on benchmark datasets commonly used in natural language processing tasks. This empirical evaluation will provide insights into the effectiveness of different approaches in capturing and quantifying uncertainty in Language Models generated text. Furthermore, the student is expected to critically reflect on how UQ can address challenges associated with Language Models, such as hallucination, misinformation, and lack of interpretability. By examining the role of UQ in mitigating these issues, the student will contribute to advancing the understanding and application of UQ techniques in the context of Language Models.
To investigate uncertainty quantification for Language Models, the project will leverage publicly available benchmark datasets [5] commonly used in natural language processing research. These datasets will provide a standardized framework for evaluating different UQ approaches across various tasks, including text generation and summarization. The selected datasets will encompass diverse linguistic phenomena and application domains, enabling comprehensive assessment of uncertainty quantification methods.
DNV is a risk management and assurance company. This means that we help our customers across various industries to make sure that their products and processes are safe, sustainable, and compliant with regulations. The Digital Assurance program in our R&D department is tasked to explore new digital technologies that DNV can use to provide better assurance or may need to assure in the future. This includes the use and assurance of AI-based systems.
[1] Chang, Y., Wang, X., Wang, J., Wu, Y., Yang, L., Zhu, K., ... & Xie, X. (2023). A survey on evaluation of large language models. ACM Transactions on Intelligent Systems and Technology.
[2] Abdar, M., Pourpanah, F., Hussain, S., Rezazadegan, D., Liu, L., Ghavamzadeh, M., ... & Nahavandi, S. (2021). A review of uncertainty quantification in deep learning: Techniques, applications and challenges. Information fusion, 76, 243-297.
[3] Abeyrathna, K. D., Hafver, A., & Edward, L. Y. (2023, August). Modeling prediction uncertainty in regression using the regression tsetlin machine. In 2023 International Symposium on the Tsetlin Machine (ISTM) (pp. 1-8). IEEE.
[4] Hwang, J., Gudumotu, C., & Ahmadnia, B. (2023, September). Uncertainty Quantification of Text Classification in a Multi-Label Setting for Risk-Sensitive Systems. In Proceedings of the 14th International Conference on Recent Advances in Natural Language Processing (pp. 541-547).
[5] Lhoest, Q., del Moral, A. V., Jernite, Y., Thakur, A., von Platen, P., Patil, S., ... & Wolf, T. (2021). Datasets: A community library for natural language processing. arXiv preprint arXiv:2109.02846.
The ability to account for uncertainty becomes important when applying AI in domains where wrong predictions can have severe consequences. Uncertainty-awareness enables the AI (or those who use it to support decisions) to apply caution when uncertainty is high in order to reduce risk. The goal is to avoid that an AI is “confidently wrong”. Many approaches exist to make AIs aware of uncertainty (e.g. Bayesian deep learning, Gaussian Process models, Monte Carlo dropout and Evidential deep learning). However, how can we know that the uncertainty expressed by such models is meaningful? If the perceived uncertainty is misguided, then an AI may still be confidently wrong or overly conservative. One step towards answering this is to investigate what determines the uncertainty of the AI.
Standard deep learning (DL) classifiers use a softmax layer for the output, which means that the probabilities for the training classes always add to 1, even if an input depicts neither of the training classes. This can lead to the classifier being confident that an image belongs to a certain class even if it shows something completely different. The student should review methods that exist for making DL classifiers uncertainty-aware and able to tell when a sample does not belong to any of the training classes. The student should implement some of these methods and test the performance when the classifiers are given images from classes not contained in the training data or manipulated images from the classes contained in the training data. The student should investigate the type and extent of manipulation (blur, contrast changes, color shifts, rotations, translations, overlays, etc.) that is necessary to confuse the classifiers and their ability to tell when a picture belongs to an unseen class or is too distorted to be classified.
Data Description:
DNV will make available an in-house dataset containing about 3000 labeled images of various types of boats. In addition, to enable comparison with other research, it is suggested that the student test out models on other openly available datasets such as RETINA benchmark and the CIFAR-10 and CIFAR-100 datasets (e.g. comparing out-of-class misclassifications between CIFAR-10 and CIFAR-100).
Challenges:It does not help that an algorithm makes good predictions in 99% of cases if the 1% bad predictions are critical. To device efficient machine learning models that can say “I don’t know” is therefore both a business- and research challenge. Techniques such as ensembling and dropout can be used to quantify uncertainty in predictions, but can impose an increased computational burden at training and test time. To adopt AI in high-risk industries, uncertainty needs to be handled efficiently. We would like algorithms that are aware of their own weaknesses and can give us an indication when they are uncertain. They should say if the input looks very different from anything in the training data (e.g. detect distributional shifts). Ideally they should quantify evidence both for and against each class in the training set. Preferably they should also be able to distinguish epistemic uncertainty (uncertainty due to lack of knowledge, which can in principle be reduced by a better model) from aleatoric uncertainty (uncertainty because of actual variation in the world, which would remain even if the model was perfect). The student is encouraged to describe these challenges and desired features in more detail and to propose and explore possible solutions.
This project is performed in cooperation with DNV.DNV is a risk management and assurance company. This means that they help their customers across various industries to make sure that their products and processes are safe, sustainable and compliant with regulations. The Future of Digital Assurance program in DNV´s R&D department is tasked to explore new digital technologies that DNV can use to provide better assurance or may need to assure in the future. This includes the use and assurance of AI-based systems.
Supervision at IDI / NTNU:Prof. Rudolf Mester
Description: Deep learning has been reported to improve upon previous state of the art in many traditional machine learning tasks, like image classification, recommender systems, text-to-speech, and so on. Nevertheless, there are still fundamentally problematic issues with these systems, that invite theoretical work on extensions of deep learning towards (traditional) probabilistic reasoning. This is a topic of some interest, which has lead to nice tools that can be used for implementation/evaluation, like Tensorflow Probability and Pyro (built on top of PyTorch). For typical research trends, see Part III of "The Deep Learning Book" by Goodfellow et al., Kevin Murphy's book-series on probabilistic machine learning, or the topics we on the program of the “ProbAI summer school”. The selection of interesting research question(s) in this area will depend on the students' interest.
Level of sophistication: This is a challenging yet rewarding topic to work on. Students should expect to do more demanding work than your run-of-the-mill project/master thesis. A good grasp of linear algebra, standard statistics, general machine learning, and deep learning is required for this to be fun. Furthermore, you will be expected to implement the ideas you develop in a deep learning framework, hence you will need to learn how to find your way around in Tensorflow, PyTorch, or similar, too.
The ongoing digital transformation of organisations and businesses along with the advancing convergence of information technology (IT) and operational technology (OT), provide significant operational challenges in terms of ensuring the stability, robustness, and security of these complex systems. As critical services become reliant on Internet-facing software systems it also makes them more vulnerable for cyberattacks. Recent studies also confirms that the number of targeted attacks is increasing, and that more sophisticated methods are taken into use.
At the same time, several voices are expressing their concern about whether current cybersecurity practices are sufficient to address these challenges [1]. A few scholars have recently started to look into how a resilience engineering perspective adopted from safety research could be used as a complementary perspective to understand the socio-technical nature of IT operations and incident response [2, 3]. This project will further study the current approaches to incident management of IT systems and aim to start untangling some of the collaborative practices that are central to sustaining software system performance.
[1] Bochman, A. The End of Cybersecurity. Harward Business Review. 2019. [2] Woods D.D. and Allspaw, J. 2020. Revealing the Critical Role of Human Performance in Software: It’s time to revise our appreciation of the human side of Internet-facing software systems. 2019. https://doi.org/10.1145/3380774.3380776 [3] Grøtan, T.O., Antonsen, S., Haavik, T.K. Cyber Resilience: A Pre-Understanding for an Abductive Research Agenda. 2022. https://doi.org/10.1007/978-3-030-85954-1_12
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.
The world beneath the waves holds a wealth of mysteries and potential discoveries. This project embarks on the journey of enhancing underwater exploration through object detection and classification techniques.
By harnessing sonar data, underwater imagery, and machine learning algorithms, the study aims to create an intelligent system capable of identifying and categorizing submerged objects with high accuracy. This research involves addressing the challenges posed by low visibility, varying environmental conditions, and diverse underwater ecosystems. The outcomes of this endeavor contribute to unlocking new insights into marine life, archaeological sites, and underwater infrastructure, ultimately aiding in the sustainable management and conservation of oceans.
With the exponential growth of online video content, extracting and analyzing spoken information from videos has become an increasingly valuable task. This thesis aims to develop an application that automates the retrieval of transcripts from YouTube videos, particularly from curated playlists, enabling the processing and analysis of multiple video transcripts at once. By leveraging large language models (LLMs), the system will categorize, summarize, and visualize trends in the aggregated data, providing meaningful insights from vast amounts of spoken content.
As a first step, students will conduct a literature review to explore existing methodologies for transcript extraction, natural language processing, and topic modeling 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:
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 project aims to investigate the efficacy and practical considerations of the use of honeypots in cybersecurity defense strategies. Honeypots can give insights into attacker methodologies and in doing so can bolster overall network security. This project will explore the practical deployment of honeypots simulating n-day vulnerable software (e.g., vulnerable software with an outstanding CVE), focusing on their ability to detect and mitigate various types of cyber threats.
The preproject will involve conducting a systematic review following the PRISMA methodology on honeypots and their application to cybersecurity. The literature review will help the candidate(s) identify research gaps and areas of interest to focus on for the MSc thesis.
Topic areas:
Sails are increasingly viewed as the most significant enhancement to the current international fleet of ships, offering a promising avenue for sustainable energy with minimal infrastructure requirements. With a rising number of vessels under development featuring sail support, and existing vessels being retrofitted with sails, the interest in harnessing wind energy is evident.
To allow wider adoption of sail ships, there is a need to have a smart planning software that accounts for optimal routing based on weather forecasting and operational requirements. This project will focus on designing user interfaces that can be used for route planning for hybrid sail ships. The proposed user interface should allow users to plan routes and oversee how the ongoing operational requirements and weather conditions may affect the planned route.
In this project we will consider how to use causal understanding to enhance a generative model, in order to obtain better generative processes. The initial idea, which of course is open for discussion, is something like the following:
Variational Auto-encoders (VAEs) are models that encode a high-dimensional datapoint (like an image) into a lower-dimensional code. Having access to the code, we can then decode this code again back to the original image. The power of this model comes from us being able to use it as a generator: Simply invent some code z ourselves, and decode that to generate a new data-point x. How then, should one make z to get an interesting object x? One well-explored solution is to ensure that the code-space is disentangled and hopefully also understandable. If one, for instance, work with a dataset of human faces, we would hope to see that one dimension of z represents the degree of which the person is smiling. We can now change the “smily-ness” of a person in a picture by encoding the image to get the code z, and simply change the code-dimension that controls smiling to suit our purpose (getting a new z') before decoding that manipulated code. This works fairly well, and has been around for some time (Example).
However, sometimes the different things that make up the code are just not independent. Say one dimension controls gender of the person depicted, another the length of the hair. If we were to intervene and change the gender of a depicted person (that is, change the gender-dimension), we may also accept or even want the system to adapt the length-of-hair accordingly. If we have such relationships stored in an extra knowledge representation, we may be able to make this happen. There are some attempts starting to explore this path (e.g., CausalVAE), and in this project we will explore these opportunities, and consider the quality of generation in light of the amount of side-information required to obtain the results.
BackgroundRecent research from NTNU’s IROS group has shown that up to 50% of radiological exams in Norway are medically unnecessary. This overuse leads to wasted healthcare resources, increased costs, and longer waiting times, delaying critical diagnostics for high-risk patients.
Hugin Medical, a Norwegian startup, is tackling this problem by using LLMs to automatically assess radiology referrals from general practitioners. The goal is to flag referrals that don't follow national guidelines, highlight missing clinical information, and assist clinicians in communicating decisions with patients. This helps reduce unnecessary imaging, frees up system capacity, and improves care for patients who need it most.
Thesis GoalIn this thesis, you will contribute to the development and validation of an AI system that evaluates radiology referrals using LLMs and other techniques. You’ll work closely with Hugin Medical and a multidisciplinary team to explore how AI can help clinicians make better, guideline-adherent decisions.
Possible MethodsYou will explore and compare different approaches to processing referral text and generating feedback letters to referring doctors:
The system should connect the referral content to relevant Norwegian clinical guidelines. While an initial prototype exists (based on synthetic data), this thesis will involve working with real, anonymized referral data from St. Olavs Hospital and Helse Midt.
What’s in it for you
Who should applyWe’re looking for:
If you are interested in this task, please reach out to Kerstin Bach to schedule a meeting to learn more about the topic.
Supervisors: Michail Giannakos
Place: LCI Lab (https://lci.idi.ntnu.no/)
Suitable for: One or two students (2 recommended)
IntroductionScience-related subjects are some of the most influential subjects that children learn at an early age due to their ability to teach them to make observations, collect data, and come to conclusions logically. Skills like these are extremely valuable in everyday life. Additionally, research shows that children’s interest in science decreases or increases between the ages of 10 and 14, depending on their learning foundation and areas of interest. Some of the most frequent arenas for supporting children’s interest in science are science centers. Visits to science centers allow children to learn about different topics and scientific phenomena. Science centers benefit from mobile and interactive technologies, but it is unclear how different elements, such as gamification and AI, can support children’s interests. With this in mind, this 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 children’s science 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 children to ask questions during their visit) in an engaging way. Based on the best practices from the literature, the candidates will develop the app (e.g., following co-design or participatory design) and then do a user study either in school settings or in a science museum 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 (M)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://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3159709
https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3159708https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3019915https://ntnuopen.ntnu.no/ntnu-xmlui/handle/11250/3019903
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 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.
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.
The long-term research objective linked to this activity is to simplify the definition and verification of coding rules, through the use of GenAI.
Virtual devices (camera, microphone) are becoming increasingly interesting for producing visual and audio effects in digital meetings. This project will investigate the creation of a virtual camera and a virtual microphone in Linux, Mac OS and Android.
Following this, some cool applications will be created, such as voice change and changing facial characteristics.
For inspiration, some well-known packages (that will not be used) include OBS Virtual Camera and VB Cable:
https://obsproject.com/forum/resources/obs-virtualcam.539/
https://vb-audio.com/Cable/
Interested in Virtual, Augmented and Mixed Reality (i.e. XR) related to learning (e.g. AI/ML/DL) and training (e.g. medical cases)
PDF version of project proposal can be found here.
Mamba architectures, based on State-Space Models, have quickly generated a lot of interest due to their potential benefits over Convolutional Neural Networks (CNNs) and Transformers. Vision Mamba (ViM) offers the potential to model long-range dependencies effectively while maintaining computational advantages compared to Vision Transformer (ViT).
These potential benefits have catalyzed substantial research endeavors in 2024, investigating the feasibility of applying ViM to the medical imaging domain, as an alternative to Convolutional Neural Networks (CNNs) and ViT.
Project Goals
This thesis project investigates Vision Mamba's potential in medical image analysis. The goals are:
The balance between theory and practice will be tailored to the student's interests.
Relevant References
Note: The field of Vision Mamba in medicine is rapidly evolving. Incorporating the most recent publications and developments will be vital for a successful thesis.
Interested? Don't hesitate to contact either of:
Since its introduction in 2017, Transformers have revolutionised NLP and completely taken over most sequence processing.
From 2020 we have also seen examples of how transformers can be used for computer vision (often called Vision Transformers) and the number of publications related to this has almost exploded in early 2021, many of which outperforms current SotA in vaiours CV tasks.
In short this project will involve two main parts: 1) the theory (i.e. really understand how transformers / vision transformers work and investigate the latest works related to transformers for CV), and 2) the practical application of vision transformers in domains like Autonomous Vehicles / Robots and Medical Image Analysis (its also posible to suggest your own project related to vision transformers). The balance between theory and practice will be found in dialog with the interested student.
NLP: Attention Is All You Need (Transformers)
CV: An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale (ViT)
You find an extensive list of Vision Transformer architectures here.
Understand and Explore the latest architectures for text to image / video / 3D for content generation (maybe in combination with AI gen. sound and music) for various applications.
Some examples:
A more concrete project, including potential application areas are decided together with students interested in the topic.
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 (e.g. u-blox) 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 potensial.
Supervisors at NTNU: Frank Lindseth and Gabriel Kiss (drop us an email if you want to know more / discuss the project)
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
The current trend in Visualization centers around APIs that can be accessed within development environments such as Python:
https://www.geeksforgeeks.org/top-python-libraries-for-data-visualization/
The objective of this project will be to compare 5 Python Visualization libraries under a task that will be defined as part of the project.
Knowledge: Python.
Sepsis is difficult to recognize and treat properly because it has symptoms like other diseases and a diverse patient population. To provide better treatments and understand different patient groups, many studies are trying to identify similar and different clinical phenotypes among sepsis patients. Clinical phenotypes include a patient’s signs, symptoms, conditions, and in-hospital events.
The objective of the master’s thesis is to develop a visual risk prototype that shows clinicians how similar their patient is compared to other patients with suspected or confirmed sepsis using clinical phenotypes. Work will be performed on medical records from St. Olavs Hospital. This project is a step towards identifying patients at risk for sepsis.
In an ongoing collaboration with Klinikk for Fysikalsk Medisin og Rehabilitering Lian St. Olav, AIT/IDI is developing a novel VR environment for designing experiments to be used in research. Currently, the main focus is to allow neuropsychologists to design experiments with automatic data collection using multiple sensors. The masters students will initially look at integrating synchronized data from an EEG braincap and an eye tracker into an existing (fairly rudimentary) VR environment, and assess the feasibility of using this as a part of the overarching goal (with regards to accuracy, ease of use, visualization of the collected data, etc.)
Furthermore, the students will expand upon the existing environment by exploring which tools are necessary to increase the usefulness of the system. This can be achieved by focusing on a phenomenon called neglect, in which a patient following a stroke ignores part of their field of view. The traditional way to assess the severity of neglect is to use pen-and-paper tests, so implementing a corresponding suite of tests in VR is a possible avenue to explore the system's usefulness while simultaneously uncovering opportunities to collect data previously unavailable using the pen-and-paper tests (for example with the use of eye tracking and/or EEG).
The findings and results from this work is intended to be built upon by other students and researchers.
In the last years, unmanned aerial vehicles (drones) have attracted the interest of both academia and industry due to their potential to change the way transportation and logistics are tackled. Drones have the potential to significantly reduce the cost, time and reliability of last-mile deliveries. In order to manage transportation by road or air or a combination of both, a vehicle routing problem (VRP) needs to be solved. The vehicle routing problem with drones (VRPD) is an extension of VRP, where drones or a tandem strategy of trucks and drones are involved in the delivery of parcels to customers. One application area of interest is the delivery of small medical packages to inaccessible, remote or dense urban areas where such packets may include blood samples, medicines and vaccines. Such VRPD need to take into account the dynamic and uncertain nature of the application area.Various masters projects are available within this topic and may include the application of various biological-inspired algorithms such as Genetic algorithsm, Multi-objective Optimisation or Particle Swarm Optimisation.
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.
Dette prosjektet gjøres i samarbeid med Institutt for matematiske fag.
For tre år siden viste en gruppe forskere innen kryptografi at det er mulig å plante bakdører i maskinlæringsmodeller. Vi søker en motivert masterstudent som ønsker å gjenskape dette resultatet og bygge videre på det, i samarbeid med biveileder og en masterstudent fra institutt for matematiske fag.
Den interesserte studenten bør først lese https://www.quantamagazine.org/cryptographers-show-how-to-hide-invisible-backdoors-in-ai-20230302/ og deretter https://arxiv.org/abs/2204.06974
Formålet med masteroppgaven er å forstå de relevante teknikkene, om bakdører kan oppdages (potensielt ved bruk av metoder fra XAI) og om bakdører kan bygges inn på andre måter.
Denne oppgaven er egnet for en motivert student med sterk bakgrunn innen programmering, stor interesse for matematikk og kryptografi, og evne til å strukturere eget arbeid og samarbeid. Det forventes at studenten samarbeider tett med en masterstudent ved MF, og aktuelle PhD-studenter ved IDI.
Interesserte studenter kan ta kontakt med marte.eggen@ntnu.no
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)
Supervisors: Frank Lindseth, Gabriel Kiss (COMP/IDI)
