Speaker: Prof. Kristin Bennett, Mathematical Science Department,
Rensselaer Polytechnic Institute

Title: Optimization Challenges in Machine Learning


Date: Thu 11 May, 2006
Time: 16:45
Place: Science Building, Room Z42

Abstract:

In this talk, we examine different strategies for capacity control in
learning and the optimization challenges that they present.  Capacity
control in learning is essential for good generalization.  Support vector
machines and related approaches use parametric penalty terms to control
capacity.  Dimensionality reduction and early stopping can successfully
control capacity as well. Beyond generalization, dimensionality reduction
also provides low-dimensional representations of the data valuable for
visualization and interpretation.  Dimensionality reduction approaches to
least-squares-loss based problems, such as PCA, frequently reduce to
generalized eigenvalue problems that are readily solved.   But
dimensionality reduction approaches for learning tasks with other loss
functions remain a challenge.  For example, how does one do dimensionality
reduction for a ranking problem? The underlying optimization problems are
typically nonconvex.  Here we will explore recent representative research
from the literature illustrating the modeling and optimization methods used
to tackle these problems.    We conclude with a novel approach for
performing dimensionality reduction for a given convex loss function and
show how it is being used for problems in cheminformatics.



Please visit http://home.ku.edu.tr/~sci-math for a schedule of
upcoming Science -Math seminars at Koc University.