Abstract:
Current prototype-based classification often leads to prototypes with overlapping semantics where several prototypes are similar to the same image parts. Also, single prototypes tend to activate highly on a mixture of semantically different image parts. This impedes interpretability since the nature of the connections between the parts is unknown. We propose a framework that is comprised of two key elements: (i) A novel method which leads to semantically coherent prototypes and (ii) an evaluation protocol which is based on part annotations and allows to quantitatively compare the explanatory capacity of prototypes from different methods. We demonstrate the viability of our framework by comparing our method to a standard prototype-based classification method and show that our method is capable of producing prototypes of superior interpretability.
