Paper
Publication
PAC-Bayes inequalities with unbounded loss functions
Abstract

We present new PAC-Bayesian generalisation bounds. The most famous PAC-Bayesian bounds are standing for classification problems, with a loss function whose range is a bounded interval, typically [0,1]. This article proposes to get rid of this condition by introducing approaches to upper-bound the exponential moment without requiring boundedness of the loss.

Authors' notes