All Optical Echo State Network Machine Learning

Conventional Artificial Intelligence (AI) technologies require a large amount of data, high energy consumption, and high computation speed. This has spawned interest in optical realizations of these networks, which use less power, have high bandwidth, and allow for faster computations. Optical reservoir computing is an emerging technology to meet these requirements and provide real time AI calculations. Many existing designs for optical reservoir computers use partially electronic components, or make use of measurement-based software activation, both of which are much slower than fully optical alternatives. Here, we propose for the first time a fully optical and hardware implementable echo state network (ESN, a subset of reservoir computing). It has a simple architecture with minimal energy and hardware requirements and performs measurement-free matrix multiplication as well as non-linear activation, as opposed to relying on slow software activation. This is done by exploiting the non-linear interaction of stimulated Brillouin scattering (SBS). The architecture is simulated assuming 10 computational nodes, and the results for its application to well-known prediction tasks are found to be comparable with a software-based ESN with the same size. These results reveal that all optical ESN is feasible with minimal hardware resources.