Kristian Woodsend (University of Edinburgh)

Kernel methods in machine learning
Tuesday 3 April 2012 at 16.00, JCMB 5327


Over the last ten years, positive definite kernel matrices have formed an important ingredient in machine learning methods. The learning process is a convex optimization in a high-dimensional feature space, yet as everything is defined in terms of kernel evaluations this high-dimensional space never has to be explicitly computed. We will follow the review paper of Hofmann, Schölkopf and Smola (sections 1 to 3, pages 1-29), covering the properties of kernel matrices, how they are constructed, and where they fit in machine learning optimization problems.

It is recommended that attendees prepare by reading the paper.

  1. Thomas Hofmann, Bernhard Schölkopf, and Alexander J. Smola, Ann. Statist. Volume 36, Number 3 (2008), 1171-1220.