Rising Stars #10 - Special Series on Conformal Prediction: Isaac Gibbs (Stanford University) 

Title: Conformal Inference With Conditional Guarantees
Abstract: Conformal inference provides a generic set of methods for transforming the point-predictions of an arbitrary black-box model (e.g. neural network, random forest) into valid prediction sets. Formally, conformal inference guarantees that its prediction sets have exact finite sample coverage marginally over the covariates. A large number of variants of conformal have been proposed that attempt to additionally obtain coverage conditional on the features. Unfortunately however, these methods provide only weak asymptotic guarantees. This is to be expected, as previous work has shown that it is impossible to obtain exact conditional coverage in finite samples. In this talk, I will discuss a family of relaxed conditional coverage objectives corresponding to coverage over classes of covariate shifts. I will then give methods for obtaining finite sample coverage under these targets either exactly, when the class of shifts is finite dimensional, or approximately when the dimension is infinite. The methods I will discuss are computationally efficient and can be easily incorporated into existing conformal inference pipelines.