The PhD defence will be in Abels utsikt - Niels Henrik Abels hus. The host of the session will moderate the technicalities while the chair of the defence will moderate the disputation.
Ex auditorio questions: the chair of the defence will invite the audience to ask questions ex auditorio at the end of the defence.
Trial lecture
January 7th 2026, time: 10:15 am, room 1259 (Abels utsikt).
Implicit bias in a classification setting
Main research findings
Large deep learning models, such as ChatGPT, are transforming society. However, there is little theoretical understanding of why these models work so well. In particular, it is unclear how they separate useful patterns from the huge amount of noise in their training data.
I researched "implicit regularization", which is the tendency for complex models to learn underlying structure without explicitly being told to do so. This phenomenon was studied in detail in two simplified models. I proved that when these models are initialized with small values, they automatically exploit the underlying structure in the data.
A better understanding of these mechanisms could enable the development of more robust, reliable and efficient AI systems.
Adjudication committee
- Assistant Professor Dominik Stöger, KU Eichstätt-Ingolstadt MIDS
- Professor Brynjulf Owren, NTNU
- Associate Professor Erik Hiltunen, University of Oslo
Supervisors
- Senior Lecturer Øyvind Ryan, University of Oslo
- Professor Anders Christian Hansen, University of Oslo
Chair of defence
Head of Department Arne Bang Huseby
Host of the session
Associate Professor Erik Hiltunen, University of Oslo