Disputation: Konstantin Holzhausen

Konstantin Holzhausen will defend his thesis "Learning Robust Representations in Neural Network Models" for the degree of Philosophiae Doctor at the University of Oslo, Faculty of Mathematics and Natural Sciences.

portrait of the candidate

 

The PhD defence and trial lecture will be streamed. The chair of the defence will moderate the disputation. 

Ex auditorio questions:  the chair of the defence will invite the audience physically present in the auditorium to ask ex auditorio questions.

→ Live streaming of trial lecture and disputation 
Due to technical issues, the disputation will begin at 9:15 AM.

→ Request for thesis copy (available until the disputation starts)

 

Trial lecture

Time and place: June 3, 2025; 6:00 PM  - Lille Fysiske auditorium (V232) - Fysikkbygningen

Title: «Local learning rules and their dynamics»

Main research findings

Understanding and controlling the behaviour of deep learning models, the most widely used approach to artificial intelligence (AI) systems, remains a major challenge. Even small changes to their input can cause unwanted behaviour. Such changes can be accidental, for example due to variations in the environment, or intentional, designed to fool the model. My dissertation asks whether the brain’s strategies for handling disturbances and uncertainty can be applied to deep learning.

To address this, I examined the mathematical structure of the internal codes neural networks use to interpret information. The first part of my thesis investigates how the statistical fingerprint of these representations affects a model’s vulnerability to imperceptible changes. The second part shifts focus to how the brain represents space for navigation, and what a reliable internal map should look like. Using mathematical models, I explored the shape and geometry of these maps in comparison to existing theories neuroscience.

Together, these findings provide insights into why the patterns we observe in the brain have certain properties, and offer mechanistic explanations for what makes some models more reliable than others. By understanding these foundations of representations, my research contributes to developing AI systems that are transparent and trustworthy.

Adjudication Committee

  • Associate Professor Il Memming Park, Champalimaud Foundation
    Portugal / Stony Brook University USA
  • Dr. Erin Grant, University College London, UK
  • Associate Professor Anis Yazidi, University of Oslo, Norway

Supervisors

  • Dr Mikkel Elle Lepperød Simula Research Laboratory, Department Numerical Analysis and Scientific Computing, Norway
  • Professor Anders Malthe-Sørenssen Department of Physics, University of Oslo, Norway

Chair of defence

Head of Department Susanne Friederike Viefers, Department of Physics, University of Oslo, Norway

Candidate contact information

LinkedIn: Konstantin Holzhausen
E-mail: holzhausen.research@gmail.com

 

 

Contact information to Department: Line Trosterud Resvold

 

Published May 20, 2025 4:56 PM - Last modified Mar. 12, 2026 8:45 AM