Using Predicates to Explain Networks

Rule-extraction from trained neural networks has previously been used to generate propositional rule-sets. While this is suucient for many applications where large rule sets can be eeectively used, it is clearly desirable to provide a more general explanation capability. In this paper we describe an approach to representing a network as a PROLOG logic program, where the activation values of hidden and output units are equated with the truth value of predicates. The technique addresses several issues. Firstly, the constructive algorithm chosen for network learning ensures hidden and output units only have activation values in f0; 1g and secondly, the decompositional rule extraction algorithm chosen produces a boolean formula for each unit, which is logically equivalent to the unit. A process is then described where the boolean formulae are translated into a rst order representation consisting of predicates, rules and facts.