Eric Moulines from Ecole Polytechnique is visiting University of Bristol and the School of Mathematics in January 2020. He will present a mini-series of lectures.
Convex optimization for machine learning
The purpose of this course is to give an introduction to convex optimization and its applications in statistical learning.
In the first part of the course, I will recall the importance of convex optimisation in statistical learning. I will briefly introduce some useful results of convex analysis. I will then analyse gradient descent algorithms for strongly convex and then convex smooth functions. I will take this opportunity to establish some results on complexity lower bounds for such problems. I will show that the gradient descent algorithm is suboptimal and does not reach the optimal possible speed of convergence. I will the present a strategy to accelerate gradient descent algorithms in order to obtain optimal speeds.
In the second part of the course, I will focus on non smooth optimisation problems. I we will introduce the proximal operator of which I will establish some essential properties. I will then study the proximal gradient algorithms and their accelerated versions.
In a third part, I will look at stochastic versions of these algorithms.
The lectures will take place at the following times:
Tuesday 28th January 11:00- 12:00
Thursday 30th January 13:00- 14:00
Friday 31st January 10:00- 11:00
As a visitor to the Heilbronn Institute he gave a series of data science lectures to Compass students on 27 November 2019
- Introduction to the variational approach and examples: Mixture models, matrix completions and recommendations, deep learning
- Theoretical analysis of variational methods
He will also present additional lectures during his visit on areas such as:
- A Generalization Bound for Online Variational Inference
Mathieu Gerber, Compass Training Co-ordinator commented: “In his lectures Pierre has provided and proved one of the first general result about the validity of variational methods, which are popular tools to approximate high-dimensional posterior distributions”