Effective differentially private deep learning