Efficient implementation of neural nets using an optimal relationship between number of patterns, input dimension and hidden nodes

Some key issues in the design of neural nets for pattern classification are topology and associated training samples required to obtain adequate performance with test samples. Currently, there does not exist an analytical framework within which to formulate the design of multilayer perceptrons. A theorem that relates input dimension, number of hidden nodes, and number of separable regions is given. The results of application to some experiments reported in the literature and to new experiments are analyzed