Integreat PhD Fellow - UiT The Arctic University of Norway, Department of Physics and Technology
Academic interests
Johan's academic interests are, among other things, explainable and interpretable deep learning models. He believes that if these models are better understood, they may be used as a proxy to gain new knowledge and insights, especially in the physical sciences. Say a models learn something valuable about gravity, and we know how to interpret and gain knowledge from the model, then we would perhaps also learn something new about gravity.
His current research is in the intersection of information theory and deep learning. He wants to figure out how information theoretic quantities may be used to gain knowledge about the learning process of deep neural networks. This is primarily done by working with the latent representation of such models.
Background
Previously Johan completed both his bachelor degree in Physics and Astronomy and his master degree in Computational Astrophysics at the University of Oslo. There, he studied the effects of General Relativity on structure formation in the early universe, attempting to find relativistic signatures. This was done by simulation large scale structure formation with both relativistic and Newtonian gravities and training deep neural networks to find the differences. This worked just partially, because it proved hard to interpret physically what these networks had differentiated between. This sparked Johan’s interest in understanding how these deep learning models work and if we can understand their learning process better.
PhD project
"On the Intersection between Information Theory and Deep Learning"
An exploration into how elements from information theory can be used to better understand deep neural networks.
Supervisors
-
Kristoffer Wickstrøm, UiT The Arctic University of Norway, Department of Physics and Technology
-
Elisabeth Wetzer, UiT The Arctic University of Norway, Department of Physics and Technology
-
Robert Jenssen, UiT The Arctic University of Norway, Department of Physics and Technology
- Ali Ramezani-Kebrya, University of Oslo, Department of Informatics