Аннотация
Despite great success in deep neural networks (DNNs) due to high accuracy, there are still major limitations (esp. lack of interpretability). In some domains such as medicine, finance, and real-time embedded systems, the interpretability problem needs to be improved because every decision needs to be clearly explained due to huge failure costs. To do this, state-of the-art works from knowledge distillation (KD) of interpretable mimic models to feed-forward DNN models for interpretability (against back-propagation based black-box methods) are adopted and compared. In this comparative study, we 1) adopt two theoretically different interpretable approaches for DNN models: a) Mimic Learning from KD methods and b) feed-forward multilayer perception (FF-MLP) using two-class LDAs, 2) reproduce and compare in terms of accuracy and interpretability using quantitative metrics, and 3) summarize the pros and cons of the two approaches with the results that mimic learning exceeds FF-MLP in the accuracy and in the number of simulatability operation counts (# SOC).