[Colloquium] Thesis Defense: Andrew Cotter

[Dawn Ellis]

Toyota Technological Institute at Chicago

In Partial Fulfillment of the Requirements for the
Degree of Doctor of Philosophy
Andrew Cotter

Will defend his thesis
Stochastic Optimization for Machine Learning

Abstract:
My thesis consists of two parts, one of which covers the application of
stochastic optimization approaches to kernel SVM optimization, and the
other to Principal Component Analysis (PCA). In my talk, I will discuss the
first of these. SVMs are an important tool because they can be used to
quickly find linear classifiers which generalize well. Using the "kernel
trick", these linear classifiers can be taken to be elements of an implicit
high-dimensional space, giving them far more flexibility and power than the
word "linear" would seem, at first, to indicate. It is important to
remember, when developing and analyzing SVM solvers, that the underlying
problem is binary classification, not merely to find good approximate
solutions to the SVM objective. I will discuss two algorithms dealing with
kernel SVMs, both of which will be analyzed not in terms of suboptimality
or some other SVM-specific metric, but rather in terms of the ultimate
quantity of interest: generalization error (i.e. classification error on
unseen data). As a result of this choice, our analysis shows these to be
useful binary classification algorithms, instead of merely being good SVM
algorithms. The first of the two, the kernelized Stochastic Batch
Perceptron (SBP), achieves the best known bound on the number of kernel
evaluations required to find a classifier with some desired level of
generalization error. The second achieves the best possible bound (up to a
small constant factor) on the support size of the resulting classifier,
which determines the number of kernel evaluations required to classify each
testing example. Combining the two yields an approach with the best known
training-time and testing-time runtime bounds on its performance.

*Date:* Friday July 19, 2013

*Time:* 10 am

**
*Place: TTIC Conference Room 526*

Faculty, students, and the general public are invited.
Thesis Advisor: Prof. Nathan Srebro

Thank you!

-- 
*Dawn Ellis*
Administrative Assistant
773-834-1757
dellis at ttic.edu

TTIC
6045 S. Kenwood Ave.
Chicago, IL. 60637
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://mailman.cs.uchicago.edu/pipermail/colloquium/attachments/20130710/d688c257/attachment.htm