Knowledge Extraction From Neural Networks : A Survey

Abstract

(eng) Artificial neural networks may learn to solve arbitrary complex problems. But knowledge acquired is hard to exhibit. Thus neural networks appear as ``black boxes'', the decisions of which can't be explained. In this survey, different techniques for knowledge extraction from neural networks are presented. Early works have shown the interest of the study of internal representations, but these studies were domain specific. Thus, authors tried to extract a more general form of knowledge, like rules of an expert system. In a more restricted field, it is also possible to extract automata from neural networks, likely to recognize a formal language. Finally, numerical information may be obtained in process modelling, and this may be of interest in industrial applications.