Evento
Descrizione:
This talk aims twofold: in its first 30 minutes, we will introduce multilayer perceptrons from scratch and their training process through gradient descent (GD).
In the remaining time, we will show how some conservation law appears by replacing the GD with a (loss) gradient
flow and how this determines invariant sets for the training. The set is a product of quadric hypersurfaces: we will
find the number of their connected component and their topological classification. If we have time, we will show how
to use the classical Segre embedding to classify these NN up to observational equivalence under a torus action.
Joint work with Marco Nurisso and Pierrick Leroy.
