- Date: 3 Dec. (Thu.)
- Place: On the Web
- Time: 18:00 - 19:30
- Speaker: Josef Teichmann (ETH Zurich)
- Title: Training algorithms and generalized Langevin dynamics
- Abstract:
We investigate generalized Langevin dynamics in the sense of Baudoin-Hairer-Teichmann and its convergence to a Gibbs measure of a loss function. We derive a pathwise decay formula for entropy as in Karatzas-Schachermayer-Tschiderer and obtain in this generalized context similar results to Hu-Ren-Siska-Szpruch, which are applied to training neural networks in machine learning.
(joint work with Robert Crowell, Christa Cuchiero, Yuuki Ida and Yuri Imamura)
(joint work with Robert Crowell, Christa Cuchiero, Yuuki Ida and Yuri Imamura)