External Seminar: Thomas Gaertner (University of Nottingham)

[Statistics & Probability Seminar]

Machine Learning Beyond Euclidean Spaces

Supervised learning problems such as regression and classification are typically studied in the context of data drawn iid from a Euclidean space. In many real-world applications, however, at least one of these assumptions is violated. For instance, data points often have an internal or external structure that permits a more natural representation as a graph, oracle feedback is available, or the range of the function to be learned is structured.

I will start my talk with a brief reminder of kernel methods as risk minimisation in reproducing kernel Hilbert spaces and then give an overview of some of my research that addresses the above mentioned issues. In particular, I will touch on (i) kernel functions for graphs, (ii) structured output prediction, and (iii) online learning on partially ordered sets and in convexity spaces. I will skip over proof details but highlight theoretical results, involved optimisation problems, and approximation algorithms.