Designing partially-connected, multilayer perceptron neural nets through information gain

An adequate number of hidden neurons and connection structure of a multi-layer perceptron network (MLP) are usually determined by experimentation. In this paper, we propose a scheme to define an appropriate structure and number of neurons of a partially connected MLP when used for classification. Rules for designing the network are based on a decision tree previously built using information gain. Our structure, called IG Net, is inspired by the Entropy Net [1], but contains fewer layers and connections than such network or than a fully-connected neural network and holds equivalent classification power. We tested the classification performance of our network using 10 databases from the UCI Machine Learning Repository. The performance obtained by IG Net using such databases showed to be statistically equivalent to the one obtained by an Entropy Net or by a fully-connected MLP, using fewer computational resources than the compared models.