Ratio Potentiality for Prototype Identification and Interpretation in Multi-Layered Neural Networks

This paper aims to show the importance of acquiring a prototype network at the beginning of learning by using new measures of potentiality and ratio potentiality. The prototype is thought to be weakly present behind complex multi-layered neural networks, making it necessary to introduce these new measures for clearer detection. Potentiality is similar to conventional entropy but is computed in a simplified and direct way. Ratio potentiality is analogous to entropic divergence but is bounded for easier interpretation. By using the ratio of actual networks to prototype ones, differences between them can be clarified, aiding in the understanding of the main inference mechanism of neural networks. The method was applied to a business dataset with three important findings. First, the potentiality could describe the learning process more clearly than the corresponding entropy. Second, the ratio potentiality indicated the possible existence of the prototype network at the beginning of learning. Third, lower ratio potentialities suggested that the number of important inputs was smaller for easier interpretation.