[BANANA] Berkeley Lab - Scientific Computing Seminar - Wednesday, December 5, 2007

[Esmond G. Ng]

Berkeley Lab Scientific Computing Seminar

Date:  Wednesday, December 5, 2007
Time:  1:00pm-2:30pm
Location:  Building 50B, 2222 Conference Room

Seminar Speaker:
   Philip Kelgemeyer
   Sandia National Laboratories, Livermore

Title:  The Counter-Intuitive Properties of Ensembles for Machine 
Learning, or, Democracy Defeats Meritocracy

Abstract:

Machine learning is the process of using past experience to predict the 
future. There are many machine learning methods; neural nets, support 
vector machines, decision trees. The design trade-offs in optimizing 
them is a tricky business, still more art than science.

"Ensembles" are a machine-learning meta-method that can be applied to 
most machine learning algorithms. Ensembles generally greatly improve 
accuracy, provably do no harm, reduce or remove most of the design 
issues, are admirably suited to parallel and distributed computation, 
and are delightfully weird and counter-intuitive.

This talk will provide an terse introduction to machine learning and 
then discuss the properties of ensembles; what they are, various 
theories on why they work, and how they can be simply applied to improve 
existing machine learning code in situ.

Sponsor of Seminar:  Xiaoye S. Li