Diminishing the number of nodes in multi-layered neural networks

We propose in this paper two ways for diminishing the size of a multilayered neural network trained to recognise French vowels. The first deals with the hidden layers: the study of the variation of the outputs of each node gives us information on its very discrimination power and then allows us to reduce the size of the network. The second involves the input nodes: by the examination of the connecting weights between the input nodes and the following hidden layer, we can determinate which features are actually relevant for our classification problem, and then eliminate the useless ones. Through the problem of recognising the French vowel /a/, we show that we can obtain a reduced structure that still can learn. >