SPEAKER:  Jorge Nocedal

ABSTRACT:

We present a "semi-stochastic" Newton method motivated by machine learning problems with very large training sets as well as by the availability of powerful distributed computing environments. The method employs sampled Hessian information to accelerate convergence and enjoys convergence guarantees. We illustrate its performance on multiclass logistic models for the speech recognition system developed at Google. An extension of the method to the sparse L1 setting as well as a complexity analysis will also be presented.  This is joint work with Will Neveitt (Google), Richard Byrd (Colorado) and Gillian Chin (Northwestern).

Short Bio:
Jorge Nocedal is a professor in the IEMS and EECS departments at Northwestern University. He obtained a BS from the National University of Mexico and a PhD from Rice University. His research interests are in optimization and scientific computing, and in their application to machine learning, computer-aided design and financial engineering. He is the author (with Steve Wright) of the book "Numerical Optimization."

He is a SIAM Fellow, an ISI Highly Cited Researcher (Mathematics Category), and was an invited speaker at the 1998 International Congress of Mathematicians. He serves in the editorial board of Mathematical Programming, and in 2011 he will become editor-in-chief of SIAM Journal on Optimization. In 1998 he was appointed Bette and Neison Harris Professor of Teaching Excellence at Northwestern.