A Method to Implement Qualitative Knowledge in Multi-Layered Neural Networks

Hiroshi Narazaki, Anca L. Ralescu

Abstract: We propose an improved synthesis method for the multi-layered neural network(NN) using a ”translation mechanism” that maps the logical representation of qualitative knowledge into a multi-layered NN structure. We give realizability conditions and synthesis equations to realize logical functions by the NN. The NN is tuned by back-propagation(BP) after the direct synthesis. This direct synthesis decreases the burden on the BP and contributes to the efficiency and accuracy of BP learning process. We demonstrate our method through function approximation and character recognition problems.