There is a sign-up schedule for this event that can be found at

http://apps.cs.utexas.edu/talkschedules/cgi/list_events.cgi

Type o

f Talk: Colloquia

Speaker/Affiliation: Dr. Lorenzo Rosasco/Massachuset

ts Institute of Technology & Istituto Italiano di Tecnologia

Talk Audi

ence: UTCS and ECE Faculty and Graduate Students

Date/Time: Thursday,
April 5, 2012, 11:00 am

Location: ACES 2.402

Host: Ambuj Tewar

i

Talk Title: Learning multiple categories with simplex coding

Ta

lk Abstract:
In this talk we discuss how a relaxation approach commonly u

sed in binary classification can be generalized to multiple classes. We s

tudy a coding/decoding strategy, namely the simplex coding, that allows

to cast multi-category classification into a vector valued regression probl

em and extend methods and results from binary classification, which is rec

overed as a special case. In particular we study and compare extensions to

multiple classes of regularized least squares and support vector machines.

We derive explicit comparison inequalities that quantify the error due to

relaxation, and highlight the differences between the two approaches. Mor

eover we discuss the computational complexity of the corresponding optimiz

ation procedures, and show that it is possible to train a consistent least
squares classifier with complexity independent to the number of classes.

Bio:
Dr. Lorenzo Rosasco is team leader of the IIT-MIT joint lab, f

ounded on a collaborative agrrement between the Istituto Italiano di Tecno

logia (IIT) and the Massachusetts Institute of Technology (MIT). He is als

o visiting scientist with the Center for Biological and Computational Lea

rning, Brain and Cognitive Sciences Dept., MIT. Dr. Rosasco received his

PhD from the University of Genova in 2006 where he worked under the supervi

sion of Alessandro Verri and Ernesto De Vito in the SLIPGURU. He was a vis

iting student at the Center for Biological and Computational Learning (CBCL

) at MIT, the Toyota Technological Institute at Chicago (TTI-Chicago) and

the Johann Radon Institute for Computational and Applied Mathematics. Betw

een 2006 and 2009 he has been a postdoctoral fellow at CBCL working with T

omaso Poggio. His research focuses on studying and developing computational
methods for modeling and analyzing complex, high dimensional system/data.
More broadly he is interested in the problem of learning, and in partic

ular in computational models of learning. Dr. Rosasco has developed and an

alyzed methods to learn from small as well as large samples of high dimensi

onal data, using analytical and probabilistic tools, within a multidiscip

linary approach drawing concepts and techniques primarily from computer sci

ence but also from statistics, engineering and applied mathematics.