Investigação sobre o efeito de ruído na generalização de redes neurais sem peso em problemas de classificação binária

Neural networks with RAM based neurons are characterized by being hardware implementable and for being well suited for solving binary-space input problems. However, for continuous-space input problems, the task of finding an adequate representation of these values without losing the classifier's generalization power proves to be quite difficult. This work investigates the use of Gaussian noise in continuous-space inputs to boost the network s generalization power. With this approach a larger quantity of memory addressable positions can be “trained”, creating neighborhoods for similar patterns, known as basins of attraction. Analyses about the influence of the addition of noise during the Boolean network training were made and they have proved that the training with Gaussian noise increases the n-tuple classifier s generalization. The performance of the investigated model was compared to results obtained by the Multi Layer Perceptron (MLP) Neural network. The experimental data for this study consisted of four public databases, three of them obtained from a well known benchmark and the fourth one from a recent international data mining competition. Experimental results showed that the investigated model performs equivalently to the MLP Neural Network for these problems.