I am a Ph.D. student here at UW-Madison.
As of May 2007 I am a member
of the Paradyn Project.
I am currently working on adapting Dyninst to make it more robust on
malware binaries that take defensive measures.
Prior to joining Paradyn, my emphasis in database research and data
mining. I looked into developing a database "Optimizer Medic" that is
activated whenever a query estimate is significantly off the actual
runtime of the query. This would result in the gathering of new
statistics or in the updating of stale histograms and other
datastructures.
It seems to me that the best classifiers are often the least
comprehensible. A classifier could be made easier to interpret by
developing hybrid learning strategies that would divide the feature
space into regions and apply simple classification models wherever
possible, while not sacrificing the overall quality of the classifier
and avoiding incomprehensible ensemble classifiers except on the most
difficult regions of the feature space. The complexity of such a
classifier is daunting, making it work at scale would be very
important.