PAC-Bayesian Learning of Linear Classifiers (2009)

Download PDF

Germain, Pascal and Lacasse, Alexandre and Laviolette, François and Marchand, Mario

We present a general PAC-Bayes theorem from which all known PAC-Bayes bounds are simply obtained as particular cases. We also propose different learning algorithms for finding linear classifiers that minimize these PAC-Bayes risk bounds. These learning algorithms are generally competitive with both AdaBoost and the SVM.