Lucas Roos M.Sc.

My research is motivated by a basic question: what makes deep learning succeed? I am interested in identifying the precise inductive biases of neural network architectures, understanding the structural properties of the data on which they are trained, and explaining how these components interact to produce effective learning. I believe image and video provide a compelling setting in which to investigate this interaction. I have some research experience in geometric deep learning, where I investigated the properties of group-equivariant CNNs, but I am also interested in exploring other areas, including compression and probabilistic modelling on images.