PAC-Bayes under potentially heavy tails

                     Matthew J. Holland
               Institute of Scientific and Industrial Research
                      Osaka University
               matthew-h@ar.sanken.osaka-u.ac.jp

                        Abstract
      We derive PAC-Bayesian learning guarantees for heavy-tailed losses, and obtain
      a novel optimal Gibbs posterior which enjoys finite-sample excess risk bounds
      at logarithmic confidence. Our core technique itself makes  ...