Extração de conhecimento simbólico de redes neurais.

Abstract

The fact that Artificial Neural Networks (ANNs) are not able to explain, in a symbolic way, neither their decisions or the knowledge embedded in its connections and architecture is a well-known limitation. This work investigates several methods of knowledge extraction from ANNs proposed in the literature. More specifically, it focuses on four different approaches for knowledge extraction that are detailed and criticized and, for each of them, discusses a possible implementation. Also, a taxonomy for methods of rule extraction from ANNs, found in the literature, is detailed. An extension of this taxonomy aiming at a more useful, refined and versatile version is proposed. The main goal of the work, however, is to approach knowledge extraction from ANN in a critical way, analyzing each of the four methods concerning, mainly, their scopes, limitations and effective contribution to improving readability and easy understanding.