We propose a new machine-learning approach for fiber-optic communication
systems whose signal propagation is governed by the nonlinear Schr\"odinger
equation (NLSE). Our main observation is that the popular split-step method
(SSM) for numerically solving the NLSE has essentially the sa