Optical next generation reservoir computing with complex media

Abstract

We propose and experimentally demonstrate an optical implementation of the “next generation reservoir computing” (NGRC) algorithm, utilizing large-scale computations via light scattering through complex media. In contrast to conventional optical reservoir computing (RC), we drive the optical reservoir directly with time-delayed inputs. We show that the optical NGRC also implicitly synthesizes polynomial features for desired functionalities, similar to digital NGRC implementation. We experimentally validate that our optical NGRC can accurately forecast both short-term and long-term dynamics of chaotic time series as well as infer unmeasured variables from observables.