MA5052 - Mathematical optimization for machine learning

Objectives

General objective:

Master the concepts and algorithms of high-dimensional nonlinear optimization based on first-order methods. The concepts and algorithms concerned are as follows:

Motivation and examples of machine learning problems
Elements of convex analysis
Gradient method and gradient method acceleration
Subgradient method
Proximal gradient method
Stochastic gradient method
Frank-Wolfe method and ADMM
Generalization to the non-convex case

Be able to put these methods into practice and analyze them. To achieve this, the course is then illustrated by 2 practical exercises, leading to the writing of a scientific article on the methods seen in class, their advantages and limitations.

Position in the programme

Linear programming, Integer programming, Simplex method, branch-and-bound, duality.

Form of assessment

Total assessment hours: 2