Efficient Neural Network Training via Subset Pretraining