I recently completed my PhD (Dec. 08) on large-scale Machine Learning.
Dissertation title: Multi-Tree Monte Carlo Methods for Fast, Scalable Machine Learning

My dissertation work focused on fast/scalable methods for massive datasets using Monte Carlo principles. For instance, we have fast Monte Carlo versions of the SVD and most kernel estimators (kernel density estimation, etc.). Our methods take only seconds or minutes to perform complex learning computations that would normally take years or decades. This enables the use of sophisticated learning algorithms on massive datasets.

Other work has included prediction in time series and dynamical systems, the Netflix prize competition (collaborative filtering), and hidden state discovery.

I was advised by Charles Isbell, and co-advised by Alex Gray.

Some papers of interest

QUIC-SVD: Fast SVD Using Cosine Trees
(NIPS 2008)
The version from my thesis shows speedups as high as 10^5 when approximating the SVD on dataset matrices as big as RAM (16GB).

Ultrafast Monte Carlo for Kernel Estimators and Generalized Statistical Summations
(NIPS 2007)
A fast training method for most kernel estimators, and for an even more general class of "nested-summative" forms that we define. Speedups as high as 10^6 on millions of points.

Fast Nonparametric Conditional Density Estimation
(UAI 2007)
A method for fast bandwidth selection in kernel conditional density estimation, which I am interested in for time series prediction. Speedups as high as 10^6 are observed.

Looping Suffix Tree-Based Inference of Partially Observable Hidden State
(ICML 2006, Distinguished Student Paper Award)
Solves the fundamental problem of inferring a purely history-based predictive state model in deterministic POMDPs. We introduce the looping suffix tree representation, which allows us to give a unified view of other hidden state methods (including schema learning) in the context of DPOMDPs.

Schema Learning: Experience-Based Construction of Predictive Action Models
(NIPS 2004)
A constructivist approach that incrementally builds up predictors of action effects.


Contact Info
mph@cc.gatech.edu
fax. 404-894-9442


© 2002-2009 Michael Holmes. All Rights Reserved.