Geometrization of deep networks for the interpretability of deep learning systems