Advanced aspects of gradient-type methods
Optimization for Machine Learning
M2 IASD/MASH, Université Paris Dauphine-PSL, 2021-2022
Program
In these lectures, we provide a modern perspective on gradient descent methods, through the prism of convergence rate analysis. We introduce the notion of momentum, as well as its applications in convex optimization. We also discuss the challenges posed in nonconvex optimization, and how recent results partially circumvent these issues.
Schedule
Lecture 1/2 (10/12) Advanced gradient descent and acceleration.
Lecture 2/2 (10/14) Nonconvex optimization and gradient descent.
Course material
Lecture notes
PDF
Handwritten notes for the first lecture
PDF
Handwritten notes for the second lecture
PDF
Python notebook for the first session
[Sources]
Python notebook for the second session
[Sources]
Materials on this page are available under Creative Commons
CC BY-NC 4.0 license.
La version française de cette page se trouve
ici.