Photonic frequency multiplexed next-generation reservoir computer

Abstract

In this work, we introduce and experimentally demonstrate a photonic frequency-multiplexed next generation reservoir computer (FM-NGRC) capable of performing real-time inference at GHz speeds. NGRCs apply a feed-forward architecture to produce a feature vector directly from the input data over a fixed number of time steps. This feature vector, analogous to the reservoir state in a conventional RC, is used to perform inference by applying a decision layer trained by linear regression. Photonic NGRC provides a flexible platform for real-time inference by forgoing the need for explicit feedback loops inherent to a physical reservoir. The FM-NGRC introduced here defines the memory structure using an optical frequency comb and dispersive fiber, while the sinusoidal response of electro-optic Mach-Zehnder interferometers controls the nonlinear transform applied to elements of the feature vector. A programmable waveshaper modulates each comb tooth independently to apply the trained decision layer weights in the analog domain. We apply the FM-NGRC to solve the benchmark nonlinear channel equalization task; after theoretically determining feature vectors that enable high-accuracy distortion compensation, we construct an FM-NGRC that generates these vectors to experimentally demonstrate real-time channel equalization at 5 GS/s with a symbol error rate of ∼ 2.5 × 1 0 − 3 .