Efficient Model Selection for Deep Neural Networks on Massively Parallel Processing Databases