Rapid Training of Very Large Ensembles of Diverse Neural Networks