ARTIFICIAL NEURAL NETWORKS FOR DATA MINING AND FEATURE EXTRACTION - Quantitative Medical Data Analysis Using Mathematical Tools and Statistical Techniques - page 322

predict the classication of an unknown data set. In such classication
problems, it is natural to ask which features of the underlying data set
are most responsible for the prediction of the classication of a data set.
Articial neural networks (ANN) can be used to address this question in
a natural and straightforward manner 3 . Although there are methods of
addressing this question with SVM's, this article will focus on the use of
ANN's as classiers which can also reveal features of the underlying data
set most responsible for that classication.
There are many dierent neural network algorithms that are used for
classication and feature extraction, including Self Organizing Machines,
the Self Organizing Tree Algorithm, perceptron networks, and multi-
layer perceptron networks (MLP) 4 . These algorithms are being used in
an ever increasing number of dierent applications. For example, articial
neural networks have been used to predict protein structures 5 , to diag-
nose lymphoma 6 , to perform clustering analyses 7 , and to interpret protein
threading scores 8 , to name a few. The list is far from exhaustive, but it
illustrates the diversity of applications of neural networks for classifying
and interpreting data.
This article describes the use of neural networks for classication and
feature extraction, with an emphasis on applications to microarray data.
The emphasis is on perceptron models, especially as they are used to clas-
sify gene expression in microarray data 3 . Section 2 introduces and explores
articial neural networks. Section 3 presents algorithms suitable for classi-
cation and feature extraction, and section 4 suggests methods for improv-
ing ANN algorithms based on mathematical models of dendritic electrical
activity.
2. Articial Neural Networks
A neuron is known to collect information in the dendrites in the form of
variations in membrane resistance and ion channel interactions at synaptic
junctions. If the resulting variation in membrane potential is not large, then
the potential decays exponentially to a resting potential of about -70 mV.
However, if the potential at the soma surpasses a certain threshold, then
the neuron \res", by which we mean that an action potential propagates
along the axon to the synapses of the neuron.
The mechanics of this process are described by the Hodgkin-Huxley
(HH) equations and by dendritic cable models with a lumped soma bound-
ary condition 10 . Because the HH equations are highly nonlinear and nearly