Title: Neural network training dynamics with symmetric data

When faced with a learning task involving an a priori known symmetry, a popular approach is to use data augmentation. That is, one transforms examples in accordance with the symmetry and trains the model on this augmented dataset. While this intuitively should bias the model towards respecting the symmetry after training, there has been an absence of theoretical guarantees. In particular, the relationship between data augmentation and so-called geometric deep learning, where the symmetries are explicitly included in the model architectures, is still unclear. This talk will present recent work on the relation between the two approaches.