MIT researchers created a mathematical framework to formally quantify and evaluate the understandability of explanations that seek to describe the behavior of a machine-learning model.