Neural Networks’ Training Software Environment with Evolutive Adjust of Topology and Activation Functions

Abstract

This research examines two problems in the optimization in the neural networks used for most real applications: first, architectural design that involves determining the number of layers and neurons by layer, and second, the activation functions that will be should use in each of these layers. For it is developed a software tool based on genetic algorithms to find these parameters of a neural network. The developed tool allows the user to choose the algorithm used for training and also apply techniques to achieve better generalization such as the early stopping, the repetition of training and adjusting the training data to the activation functions used. Finally, the developed tool is tested into a specialized group of users who use the tool to find an optimal neural network architecture to solve a problem of identity verification through the facial image using artificial neural networks.