Designing neural network explanation facilities using genetic algorithms

The authors describe the use of genetic algorithms to provide components of explanation facilities for neural network applications. The genetic algorithm implementation, Genesis, uses a trained backpropagation neural network weight matrix as the genetic algorithm fitness function. Using different combinations of Genesis´ run-time options, codebook vectors and decision surfaces are defined for the trained neural network. These vectors and surfaces can then be used as components of a facility that explains how the network is trained, and how it differentiates between classes. Two examples of this methodology are presented and briefly discussed. The first is a network trained to solve the XOR problem. The second is a network trained to diagnose appendicitis