Artificial intelligence (AI) in the Department of Computing Science’s research.
Artificial intelligence (AI) research at Umeå University is shaped by a vibrant community of researchers who address fundamental challenges in developing new theory, methods and techniques to advance AI, exploring how intelligent systems can benefit society, environment and humanity. Umeå University's AI research addresses a wide spectrum of AI, from advancing machine learning, optimization, and natural language technologies, to understanding AI-supported human decision‑making and building trustworthy autonomous systems that are reliable, transparent, sustainable, while grounded in current societal needs.
Through collaborations across research domains and societal organisations new research avenues are explored where AI can be further utilised as instrument to support people in making better decisions, improving health and wellbeing, and address fundamental research challenges in areas such as life science, medicine, social work, education, law and regulation, safety and sustainability.
Feras M. Awaysheh : Collaborative and Privacy-Preserving Intelligent Systems
Research Description
My research focuses on the design and development of intelligent systems that can learn, scale, and collaborate across distributed environments while preserving privacy, security, and trust. I investigate federated learning, edge intelligence, and distributed AI techniques that enable parties to jointly build AI models without sharing sensitive data. A key aspect of my work is addressing challenges related to data heterogeneity, resource constraints, communication efficiency, and trustworthy AI deployment in real-world environments. The overall goal is to enable scalable, efficient, and responsible AI systems for applications ranging from healthcare and cybersecurity to smart cities and industrial automation.
Keywords
federated learning, edge intelligence, distributed and trustworthy AI, privacy-preserving analytics
My research interests centre on hybrid approaches to representing and processing multimodal data, combining symbolic and neural methods at both foundational and applied levels. I am equally engaged in principled, accountable evaluation of generative AI, and co-founded the Measuring What Matters initiative, which aims to raise the quality and impact of evaluation practices. Beyond research, I serve as program manager for WARA Media and Language and chair the Scientific Team of the SELLMA LLM project.
Keywords
LLMs, multimodal systems, evaluation, graph-structured data
Timotheus leads the Agent and Reasoning group. One of his focus areas lies in the intersection of processes and genAI agents that currently disrupt scientific workflows. Timotheus has substantial experience in data science and data-driven AI, most notably as Principal Scientist at SAP, Europe's largest software vendor. The Agents and Reasoning group frequently collaborates with researchers across disciplines on AI projects, as well as with industry. Notable past and current collaboration partners are at organizations such at IBM, SAP, Ercisson, Oxford University, the Technical University of Munich, and Imperial College London.
Helena Lindgren : AI-based assessment/decision-support/intervention studies in the medical and health domains
Research Description
AI-based assessment/decision-support/intervention studies in the medical and health domains
We design, develop and evaluate AI-based systems in collaboration with researchers in the medical and health domains, exploring how agentic AI and other kinds of interactive AI systems should be designed to support a person’s learning, reasoning, decision-making in the context in which it is aimed to be used. The representation and utilisation of domain knowledge particular for a situation is instrumental. Application examples are decision-support system for dementia diagnosis, assessing risk for occupational-related injuries for people in the mining and construction industries, person-tailored agent-based support for behavior change to improve health and prevent cardio-vascular disease and exhaustion syndrome.
Keywords
knowledge representation and reasoning, machine learning, human-AI collaboration and teaming, user modeling and personalisation, agentic AI, participatory design
Research Description
AI for identifying phenotypes that affect plasticity and learning in humans
We explore how rich data sets collected in studies such as the Västerbotten Intervention Program (VIP) can inform the design and implementation of person-tailored interventions aimed at improving learning in humans. Trajectories of change across time, as well as particular phenotypes could be identified, used for measuring effects of interventions as well as for informing the person about progress and health.
Keywords
neurology, brain plasticity, human learning, machine learning, causal reasoning
Research Description
Agentic AI for human-AI teaming in medicine and health
We explore how agentic AI can be developed and utilised in particular situations to support collaborative tasks and teamwork, including support for diary studies in the rehabilitation and health domains. We explored the concept and application of the virtual physiotherapist in a falls prevention intervention, and currently the Virtual Occupational Therapist (VOT), to assist data collection in diary studies with older adults. Application domains are dementia diagnosis and supporting individuals in their pursue of lifestyle changes to improve health.
Keywords
large language models, generative AI, human-AI teaming, digital health coaching, multi-agent systems, diary studies
Vicenç Torra: Privacy-preserving machine learning / Approximate reasoning and decision making
Research Description
Privacy-preserving machine learning
We study privacy technologies on the whole pipeline from data until machine learning models. This includes privacy models (as k-anonymity, differential and integral privacy), privacy methods, high quality privacy-preserving synthetic data generators, and disclosure risk measures (including different types of attacks).
Keywords
data privacy, privacy-preserving machine learning, disclosure risk, membership inference attacks, synthetic data generators
Research Description
AI security
Research focuses on aspects related to information leakage, membership inference attacks, and recently we are working on topics related to memory poisoning attacks. For each of these topics we are also interested in mitigation strategies.
Keywords
information leakage, membership inference attacks, memory poisoning.
Research Description
Approximate reasoning and decision making
Research focuses on mathematical formalisms to model reasoning and decision making under uncertainty. We focus on tools related to fuzzy sets and non-additive measures and integrals.
Keywords
fuzzy sets and systems, non-additive measures and integrals, multi-criteria decision making, game theory
I have access to a thorough pipeline for using AI methods for qualitative research, including: transcribing interviews or recording; analyzing qualitative data, including inductive analyses (via e.g., topic modelling) and deductive analyses (via natural language classification) — all in compliance with GDPR. This pipeline allows reducing the time for performing qualitative coding as well as scaling up qualitative coding to larger datasets, thus enabling statistical validation and exhaustive used of data. Specialized qualitative analyses can be carried out on demand for specific issues (e.g., non-verbal communication).
Keywords
transcription, qualitative research, inductive analyses, deductive analyses, large qualitative analyses
