14/03/2024, 10:00 — 10:15 — Room P3.10, Mathematics Building Online
Gonçalo Oliveira, Instituto Superior Técnico
Phase transitions and the training of neural networks
In recent years mathematics and physics-inspired ideas have been prominent in an attempt to understand how neural networks work. The design and training of neural networks depend on several parameters and analyzing how macroscopic properties of the network change concerning variations of such parameters leads to transitions and/or bifurcations in behavior. These can be explicitly analyzed from a rigorous mathematical perspective leading to a better understanding of the phases on which networks learn best.
Pre-requisites: calculus and linear algebra.
See also
10H00_GO_h.pdf
