Abstract
In this chapter we explore the generalization power of perturbation models. Learning parameters that minimize the expected task loss of perturbation models amounts to minimizing PAC-Bayesian generalization bounds. We provide an elementary derivation of PAC-Bayesian generalization bounds, while focusing on their Bayesian components, namely their predictive probabilities and their posterior distributions.
Original language | English |
---|---|
Title of host publication | Perturbations, Optimization, and Statistics |
Editors | T Hazan, G Papandreou, D Tarlow |
Pages | 289-309 |
Number of pages | 21 |
DOIs | |
State | Published - 1 Dec 2016 |