Nonlinear random matrix theory for deep learning

  Abstract