Nonlinear Advantage: Trained Networks Might Not Be As Complex as You Think