Treinamento de redes neurais com incorporação da técnica Backpropagation ao FDIPA

Abstract

Artificial Neural Networks are mathematical and computational models inspired by the functioning of the human brain. They are able to learn and perform tasks such as pattern recognition, image classification, credit card fraud detection and other applications. The structure of a network is composed of nodes (which are the neurons) connected by edges (which are the conections) distributed in layers. These conections have values (weights) that represent how important the connection is to the determination of the final result. The computation of the network is given by a series of composition of functions (activation functions) applied to the product of the weights by the values attributed to the neurons in each layer. In order to the network to learn, optimization techniques must be applied to determine the optimal network weights. The main objective of this work is to incorporate the backpropagation technique to the optimization algorithm FDIPA - Feasible Directions Interior Point Algorithm to obtain the optimal weights of a neural network. After completing this task, several tests were carried out to prove the efficiency of the proposal.