Geometric deep learning seminar / half-way seminar

Abstract: Geometric Deep Learning (GDL) is a vast and rapidly advancing field. In this talk, I provide a brief introduction to GDL, some approaches and applications, along with a few examples. An essential aspect of GDL is how the model handles symmetries in data or spaces. For data defined on a manifold, one such symmetry is the choice of local coordinates and I will present our formulation of a convolutional layer which is equivariant to the choice of local coordinates (gauge equivariant) along with a brief overview of the required structures and concepts. Finally I will discuss our findings on the benefits and drawbacks of enforcing equivariance in the model compared to augmenting the training data. Based on our findings I am going to present some questions to consider when choosing the best approach for your models.