Learning by Ignoring, with Application to Domain Adaptation