Reliable Stochastic Optimization

Aaron Defazio (ANU / NICTA)

NICTA SML SEMINAR

DATE: 2013-07-18
TIME: 11:15:00 - 12:15:00
LOCATION: NICTA - 7 London Circuit
CONTACT: JavaScript must be enabled to display this email address.

ABSTRACT:
Stochastic optimization methods are generally considered superior to batch optimization methods for most machine learning applications, however they require considerable tuning in order to work at all. In this talk I will discuss a recently published method from INRIA that does not require any tunable parameters. It is applicable to the strongly convex, non-online case only. It has some remarkable properties that are not yet fully understood.
BIO:
http://users.cecs.anu.edu.au/~adefazio/