Thus many variants of the LDA procedure are suitable for network
construction. Empirical comparison of the simplest version based on
pseudoinverse solution for the Hepatitis dataset [12] gives 86.4% accuracy
(10-fold crossvalidation tests) versus 82.1% for randomly initialized
MLP, for the Cleveland Heart disease 84.5% versus 81.3%
and for the Pima Indian Diabetes 77.5% versus 76.4%. For the Wisconsin
breast cancer data LDA accuracy is 96.0%, while MLP after
training achieves slightly better result of 96.7%. In all cases networks
with a single neuron are constructed since these are two-class
problems. We are experimenting now with adding more than one
hyperplane to increase the complexity (and hopefully the accuracy)
of MLP networks.