Learning from a lot: Empirical Bayes in high-dimensional prediction settings