Learning Theory Circle · Sunday, Aug 16: Mahdi Mostafaei on Why SGD prefers flat minima

[Learning Theory Circle]

Hi all,

Today's Learning Theory Circle session:

Speaker:  M.Mahdi Mostafaei (Shahid Beheshti University)
Paper:  Why Does SGD Prefer Flat Minima? Through the Lens of Dynamical Systems
Paper link:  https://openreview.net/forum?id=Zffca0v-i5S
When:     Sunday, August 16 · 7:00 PM Tehran (UTC+3:30)
Where:    https://meet.google.com/csa-sbkw-ctm

Abstract

The paper shows that stochastic gradient descent (SGD) escapes from sharp minima exponentially fast even before SGD reaches a stationary distribution.

Everyone is welcome — no preparation needed, and questions are encouraged
throughout.

Full schedule and past talks: https://learning-theory-circle.github.io/
Would you like to present a future session? Just reply to this email.

See you,
Erfan