Estudo do uso da computação paralela no aprimoramento do treinamento de redes neurais artificiais feedforward multicamada

Abstract

This work compares classical feedforward neural networks with an algorithm that permit exploit parallelism present in widely available machines. The training scheme discussed here do not depend on the neural network topology, therefore may be a better representation of the natural neural networks produced along animal evolution. Although, for being a strategy that does not use derivative of the activation function, is more rustic, that may result in a slower convergence when compared to topology dependent algorithm, like Error Backpropagation. An improvement in the convergence speed is expected with the use of parallel computing. However, in future works, parallelism in machines with many cores may more than compensate this lower speed, once the weights are ajusted independently. The behavior of the most popular training method nowadays, the Error Backpropagation, and the algorithm discussed in this work, in either sequencial and parallel mode, will be analised and compared.