A knowledge representation formalism generated from feedforward neural network

The present work develops a connectionist knowledge acquisition tool, enabling the user to understand the behaviour of data sets and neural networks used. In this paper, a methodology is presented which trains a feedforward neural network, prunes out the redundant links and irrelevant attributes from the network and then extracts knowledge in disjunctive normal form (DNF). The propositional (i.e. variable free) rules are processed into a representation formalism based on the rules with variables and n-ary predicates, facts and type hierarchy. We applied SHRUTI reasoning system to search the generated knowledge base (equivalent to the neural network) for events of interests to explain (how and why) the decision process of neural network. The empirical results show that`rules with variables and n-ary predicates' can be obtained by the connectionist approach with high accuracy and a knowledge base is successfully automated to reason with any inference tool.