Trial lecture
“From Convolutions to Transformers: How Modern Vision Models See Images”
Time and place: February 10, 2026 11:15 AM. Kristen Nygaards sal (5370), Ole-Johan Dahls hus.
Main research findings
Today’s AI vision models can recognize faces, animals, and objects with astonishing accuracy, and they keep improving as we feed them more data and provide them with more computing power. But this progress comes at a price: modern vision systems are increasingly large, energy-hungry, and difficult to interpret or understand.
This thesis argues that the problem is not only how much data machines see, but how they see it. Humans do not analyze images pixel by pixel; instead, we focus on shapes, regions, and relationships in visual scenes. The research presented here shows that artificial intelligence can be convinced to do the same. By letting models learn how images naturally divide into meaningful visual regions, they can focus on structure and form rather than minute pixel-level details.
The result is a new way of building vision systems that are more effective, accurate, and interpretable. As visual AI becomes part of everyday life it becomes increasingly important to develop models that are faster, more transparent, and more sustainable.
Adjudication committee
- Associate Professor Arno Solin, Aalto University, Finland
- Professor Kerstin Bach, NTNU, Norway
- Adjunct Professor Roy Edgar Hansen, Department of Informatics, University of Oslo, Norway
Supervisors
Professor Gerberth Adin Ramirez Rivera, University of Oslo, Norway
Professor Anne Helene Schistad Solberg, University of Oslo, Norway
Odd Kolbjørnsen, University of Oslo, Norway
Chair of defence
Contact information at Department: Pernille Adine Nordby