There used to be a lot of research into using deep NNs to train decision trees, which are themselves much less of a black box and can actually be reasoned about. I wonder where that all went?