Adaptive norms for deep learning with regularized Newton methods