Post doctor (2 years) position in a WASP-financed project at the The Department of Mathematics and Mathematical Statistic at Umeå University.
Your work assignments
Geometric deep learning refers to the study of machine learning problems involving a-priori known symmetries. A simple example are convolutional neural networks, that are known to be invariant towards translations. Using representation theory, one can build networks that are invariant, or more generally equivariant, towards a very broad class of symmetries. These networks are mathematically highly intriguing, but have also been successfully used in several science-related applications.
Explicitly accounting for symmetry is not the standard approach in machine learning – a well set up model should be able to ‘detect’ the symmetry automatically. One way to encourage the model to respect the symmetry is to explicitly make sure that the data is symmetric – one speaks about data augmentation.
In order to understand the relation between these approaches, we need to understand the effect of symmetric data on the training of neural networks. This will be the overarching goal of the post doc. The problem can be approached from many different angles. We are mainly interested in studying the problem theoretically.
The chosen candidate will be given the opportunity to work within the group for mathematical foundations for AI. This young, dynamic and international group is an environment bringing excellent opportunities for the researcher to develop their scientific interests and strengthen their qualifications. They are expected to work together with the current researcher in the groups on existing projects, but will also be given the opportunity to develop and work on their own research ideas.