“Zero-pretraining deep learning methods are achieving competitive performance with networks as small as 7 million parameters.”