Enhanced Rule Extraction by augmenting internal representation of feed forward neural network

Abstract

Human readable symbols are extracted from a trained neural network using Rule Extraction Techniques. In this paper internal representation of feed forward neural network is augmented by a distance term to produce fewer rules. This paper presents an efficient method to extract fewer rules from multilayer feed forward neural network. The proposed method calculates distance between activation values of hidden units for a given input values and moves them depending on the calculated distance value. The method shows fewer rules on three publicly available data sets without compromising classification accuracy.