Learning Approximately Objective Priors