Topological Interpretability for Deep-Learning