Ryan Riegel

The below is presently outdated, and I apologize. Anyway, publications:

SDM 2008: Massive-Scale Kernel Discriminant Analysis: Mining for Quasars

Compstat 2006: Large-Scale Kernel Discriminant Analysis with Application to Quasar Discovery

One of these days I'll make a more exciting webpage, with pictures and other cool things. But for now, this should suffice.

Personal Information

I'm a second-year Ph.D. student associated with both ISS and CSE. I obtained a BS in Joint Mathematics and Computer Science from Harvey Mudd College in May 2005 and have no master's degree. I am working with Prof. Alex Gray and my primary focus is computational statistics, which consists of finding quick, novel, or otherwise interesting code solutions for the analysis of data, and is especially concerned with massive data sets or high dimensionality.

Schedule

I was fixing to put this into a snazzy table, but then I remembered that this is html, not LaTeX. Along with the general theme of this site, I'll get around to makeing this look nicer in a bit, for relative sizes of "bit."

Actually, I should just link to my

In general, I'll arrive at the TSRB sometime before my scheduled seminars there, or noonish on days I don't have scheduled obligations, and I don't leave until 9:00 or so unless I'm being particularly unproductive. I'm there pretty much everyday of the week and some weekends, too. My desk is near Alex Gray's office.

Projects:

Algorithmica

While one might say that this project hasn't officially started yet, my research group and I are at least thinking about it rather intently. Algorithmica is intended to be a system capable of automatically deriving and implementing algorithms to solve user-specified problems. Its basic method will be to stitch together highly general, pre-programmed code segments that address certain computational tasks in a manner it deduces to be both effective and efficient for the problem at hand. This project is based on previous work under the name AutoBayes by Bernd Fischer, Johann Schumann, Wray Buntine, and Alex Gray.

Implementing applications of Generalized N-body Problems

The above project will ultimately be able to reproduce a great deal of the work encompassed here, but in the meantime, my efforts are directed towards the class of problems known as Generalized N-body Problems. These are characterized as computational problems that involve comparisons between groups of data points that exist in a metric space, and can often be solved efficiently with simultaneous expansion upon pairs (or larger groups) of spatially-informed trees. Notable examples are all-nearest-neighbors, kernel density estimation (KDE), non-parametric Bayesian classification, the n-point correlation, and many others. Currently, I am working on improving the speed of SMO (the state-of-the-art in training support vector machines) through the application of our methods.

Investigation of L2E/L1E as a general objective function

Minimizing integrated square error (L2E) is a common obective in non-parametric problems (e.g., KDE), but it can be used for parametric fitting as well. It demonstrates increased resistance to outliers and other noise over maximum likelihood (ML), which is the defacto objective function of EM algorithms. Further, integrated absolute error (L1E) may offer even further robustness, though until now, it has not been considered to be a feasible objective function due to its inconvenient mathematical properties. It is, however, possible to overcome L1E's difficulties with an EM-like surrogate optimzation that converges to L1E's results while only performing L2E-like work.

Contact Information:

rriegel -at- cc.gatech.edu

College of Computing

Georgia Institute of Technology

Atlanta, GA 30332-0280