https://blog.fastforwardlabs.com/2017/08/02/new-research-on-interpretability.html