Nonlinear random matrix theory for deep learning

   Abstract