Tue, 28 Apr 2015
14:00 -
15:00
L3
Newton-type methods for Support Vector Machines and Signal Reconstruction Problems
Kimon Fountoulakis
(University of Edinburgh)
Abstract
Support vector machines and signal reconstruction problems have initiated a resurgence of optimization methods with inexpensive iterations, namely first-order methods. The efficiency of first-order methods has been shown for several well conditioned instances. However, their practical convergence might be slow on ill-conditioned/pathological instances.
In this talk we will consider Newton-type methods, which aim to exploit the trade-off between inexpensive iterations and robustness. Two methods will be presented, a robust block coordinate descent method and a primal-dual Newton conjugate gradients method. We will discuss theoretical properties of the methods and we will present numerical experiments on large scale l1-regularized logistic regression and total variation problems.
