The functional localization of neural networks using genetic algorithms.

Abstract

We presented an algorithm for extracting Boolean functions (propositions, rules) from the units in trained neural networks. The extracted Boolean functions make the hidden units understandable. However, in some cases, the extracted Boolean functions are complicated, and so are not understandable, which means that the hidden units are not functionally localized. This paper presents an algorithm for the functional localization of (the hidden units of) neural networks. When a hidden unit is well approximated to a low-order Boolean function, the unit can be regarded as functionally localized. The functional localization of a hidden unit is evaluated by the error between the hidden unit and the low-order Boolean function extracted from the hidden unit. The optimization is executed by genetic algorithms. We applied it to vote data, mushroom data and chess data. Experimental results show that the algorithm works well.