Uso de conhecimento preliminar na melhoria do aprendizado em um modelo simbólico-conexionista.

Abstract

The Combinatorial Neural Model (MNC) is a neural network of superior order that is adequate to classification tasks. The training of this network is performed by an algorithm based on backpropagation, through punishments and rewards, according to the examples of the training set. This work detaches the importance of the background knowledge to learning process in a general form and describes how semantic relevance and attribute valoration can improve both the quality of learning accomplished by MNC and the performance of the training algorithm. The presented alterations actuate on the topology of the network to be trained and are compatible with the extensions that have been proposed to the Model.