Leveraging nonlinear dynamics in silicon microring resonator arrays for image classification via reservoir computing

We study the use of a neural network on the silicon photonics platform consisting of a network of coupled microring resonators and featuring multiple optical ports. This circuit exhibits recurrent nonlinear dynamics and short- and long-term plasticity due to the nonlinear effects in silicon, namely based on free carriers absorption and thermo-optic effects. Our neuromorphic hardware was devised to improve machine learning on optical signals, in terms of energy efficiency, speed and applicability. It is employed without the requirement of knowing its internal dynamics and can produce multiple different nonlinear representations of input data through different output ports and laser frequencies. Thus it provides a powerful platform for edge computing based on the reservoir computing approach. As a proof of concept, this network is used for handwritten digits classification. By adjusting the power level in the background of the images, we exploit the nonlinear dynamics and the network memory to expand data dimensionality, which significantly enhance the reservoir’s overall performance.