Knowledge extraction from neural networks for signal interpretation

摘要

Artificial neural networks have proved their ability to perform classification tasks. This ability is not satisfactory when expertise of the application domain is not available or when experts want to know more about hints that led to the decision. This leads presently to a great amount of work for knowledge or rule extraction from neural networks. In this paper, we propose a technique able to extract rules and to explain the functioning of the hidden layers of a multilayer perceptron. The first step consists in pruning the network with the classical OBD algorithm. Then, tightening of the sigmoidal transfer function can simply result in such knowledge extraction. This principle has been first tested on an application of signal interpretation in the radar domain.