A Survey of PAC-Bayesian Learning Theory for Big Data

The theory of probably approximately correct(PAC)is a framework for the study of learnable.In recent years,researchers combined Bayesian method with distribution-free PAC guarantees and proposed so-called PAC-Bayesian learning theory.This theory can give generalization error bounds for an arbitrany prior measure on an arbitrary concept space,so it has been widely used in different fields of artificial intelligence to analyze related algorithms.This paper surveys the derivation of PAC-Bayesian learning theory and its core ideas.Further,considering the characteristics of big data,this paper discusses why PAC-Bayesian is useful for theoretical analysis of the related algorithms for big data.