In this talk we introduce conformal prediction, a distribution-free approach to constructing prediction sets for risk control. We describe its implementation, discuss its advantages and disadvantages, and provide its applications to prediction and classification problems with real-world and simulated examples. We also briefly discuss "Learn then Test'', a newly-developed risk control methodology that can enhance performance of conformal prediction on a wide array of machine learning tasks.         

Keywords: covariate shift; deep learning; empirical distribution; multiple hypothesis testing; quantile regression.