Title: Introduction to conformal prediction
Abstract: Bayesian statistics provides ideal predictions for future observations provided its assumptions are satisfied. Conformal prediction can be regarded as a way of making Bayesian predictions (or predictions that depend on a parametric statistical model) more robust. The validity of conformal predictions only depends on the observations being IID, which is a standard assumption in machine learning and nonparametric statistics. If conformal predictions are packaged as set predictions, their validity means a guaranteed coverage probability. If they are packaged as predictive distributions, it means calibration in probability. In this talk I will discuss both validity and efficiency of conformal predictors. If time allows, I will also mention current areas of research, such as testing the IID assumption using conformal prediction.