Statistical Learning Theory for Modern Machine Learning

The talk introduces the motivation behind statistical learning theory and after reviewing some earlier approaches gives an overview of the PAC-Bayes methodology to analysing generalisation, highlighting the links between Bayesian and frequentist ideas that it embodies. Its application to Support Vector Machines and Deep Neural Networks will be discussed in more detail demonstrating its value in assessing performance and motivating different algorithms, but also exposing the mismatches between theory and practice.