Towards Scalable, Private and Practical Deep Learning