A UdeM-McGill-MITACS machine learning seminar will be held this Thursday, July 15th. The talk given by Ludovic Arnold, will take place from 13h30-14h30 in the room AA1175 (pavillon Andre-Aisenstadt, click herehttp://maps.google.ca/maps?q=2920,+chemin+de+la+tour&hl=en&z=16&iwloc=Afor directions) at the Université de Montréal. Hope to see you there !

Title: Information-Geometric Optimization Algorithms: A Unifying Picture via Invariance Principles

Abstract: In the context of gradient descent for likelihood maximization, the choice of a metric is critical. By using a so-called "natural gradient descent" strategy in which a meaningful metric is chosen, one can benefit from agreeable properties such as invariance w.r.t. parametrization and faster convergence. After an introduction to natural gradient descent in the canonical case of likelihood maximization, I will present a unifying framework for Information Geometric Optimization which gives a new understanding of Evolution Strategies such as CMA-ES, NES, CEM and PBIL.