PhD Thesis on Mixtures of Experts Available

[Steve Waterhouse]

Dear Connectionists,

I am pleased to announce the availability of my PhD thesis,
"Classification and Regression using Mixtures of Experts" which is now
available on my web site:

http://www.ultimode.com/stevew/thesis.html

An outline follows:

Classification and Regression using Mixtures of Experts
by Steven Richard Waterhouse October 1997,
Cambridge University Engineering Dept.
and
Jesus College Cambridge


This thesis investigates a recent tool in statistical analysis: the
mixtures-of-experts model for classification and regression. The aim
of the thesis is to place mixtures-of-experts models in context with
other statistical models. The hope of doing this is that we may better
understand their advantages and disadvantages over other models. The
thesis first considers mixtures-of-experts models from a theoretical
perspective and compares them with other models such as trees,
switching regression models and modular networks. Two extensions of
the mixtures-of-experts model are then proposed.  The first extension
is a constructive algorithm for learning model architecture and
parameters which is inspired by recursive partitioning. The second
extension uses Bayesian methods for learning the parameters of the
model. These extensions are compared empirically with the standard
mixtures-of-experts model and with other statistical models on small
to medium sized data sets. In the second part of the thesis the
mixtures-of-experts framework is applied to acoustic modelling within
a large vocabulary speech recognition system. The mixtures-of-experts
is shown to give an advantage over standard single neural network
approaches on this task. The results of both of these sets of
comparisons indicate that mixtures-of-experts models are competitive
with other state-of-the-art statistical models.


-------------------------------------------------
Steve Waterhouse
Ultimode Systems
Data Mining Tools

2560 Bancroft Way #213,
Berkeley, CA 94704

T: 510 548 8978 F: 510 845 2292 M: 415 420 3365
E: stevew at ultimode.com W: http://www.ultimode.com
-------------------------------------------------