Speaker
Piotr Korcyl
(Jagiellonian University)
Description
Gradient flow equations provide continuous deformations of fields (gauge potentials, metrics, neural weights, etc.). They have been extensively studied and used in classical gravity, where they have allowed solving important problems. They are also employed routinely in lattice QCD calculations as a tool to define the strong coupling constant and thus to set the physical scale of simulations. The gradient descent algorithm used to train neural networks is yet another version of the same gradient flow equation. In the talk, I will introduce the gradient flow and describe its advantages and relationships in various contexts.
Author
Piotr Korcyl
(Jagiellonian University)