Workflow for Knowledge Extraction from Neural Network Classifiers

Artificial neural network classifiers are widespread models used by many machine learning engineers. Although due to fact they are black box models, in mission critical areas (like healthcare, finance, atomic power) when explainability is required they cannot be used even when they show higher classification performance in comparison to explainable models like decision trees. To mitigate this problem knowledge extraction algorithms have been proposed allowing to extract knowledge in different forms. Current paper gives a review of three knowledge extraction algorithms, presents their strengths and weaknesses. Finally knowledge extraction workflow utilizing abovementioned algorithms is described.