Tuning the distribution dependent prior in the PAC-Bayes framework based on empirical data

In this paper we further develop the idea that the PAC-Bayes prior can be defined based on the data-generating distribution. In particular, following Catoni [1], we refine some recent generalisation bounds on the risk of the Gibbs Classifier, when the prior is defined in terms of the data generating distribution, and the posterior is defined in terms of the observed one. Moreover we show that the prior and the posterior distributions can be tuned based on the observed samples without worsening the convergence rate of the bounds and with a marginal impact on their constants.