This paper introduces a novel, computationally-efficient algorithm for
predictive inference (PI) that requires no distributional assumptions on the
data and can be computed faster than existing bootstrap-type methods for neural
networks. Specifically, if there are $n$ training samples,