Nonlinear polarization effects in optical fiber-based extreme learning machines

Extreme learning machines (ELMs) based on nonlinear propagation in optical fiber are a novel platform for unconventional computing. However, although experiments are showing promising results, there has been less work developing realistic models to study how such systems might exploit different experimental degrees of freedom. Here, we use numerical simulations of the coupled vectorial generalized nonlinear Schrödinger equations to present the first, to the best of our knowledge, study of fiber ELMs including polarization effects during propagation. Polarization evolution is found to significantly influence performance, necessitating care in polarization selection during readout. We also show that concatenating multiple polarization states during readout yields superior results compared to the analysis of a single state.