Digit classification using photonic reservoir computing based on a silicon microring resonator

Abstract

We demonstrate first experimental investigation on the performance of a single-node reservoir computer based on a silicon microring resonator (MRR) operating on the digit recognition task. The input layer of the reservoir is composed of a single laser, a Mach-Zehnder electro-optic modulator, which encodes intensity of the light applied to the MRR input. The input signal is transformed into a virtual high-dimensional space through thermal nonlinearity in the MRR. The MRR response is recorded with readout network consisting of a photodetector and an oscilloscope. To study the principle of operation we measure nonlinear frequency response as well as dynamic response of the MRR. The resonator demonstrates a negative shift of the resonance frequency with an increase in input power due to the dominating thermo-optic effect. In addition to the frequency shift, the MRR transmission coefficient grows at the red side of the resonance. This effect underlies the nonlinear transient dynamics at the MRR output and provides an intrinsic fading memory that are the basis for the implementation of a reservoir computer. Here a silicon MRR serves as a single nonlinear physical node. We give proof of concept demonstrations of the developed reservoir architecture by solving the classification task. The performance characteristics is evaluated with the short-term memory and the parity-check tests. Obtained results pave the way to chip-scale optical reservoirs computing.