In this work, we propose a novel design for optical neural network system based on reservoir computing. The optical neurons in the system consist of directional couplers, optical fibers, and optical amplifiers. The nonlinear activation function of optical neuron can be obtained by the optical amplifiers. Mach–Zehnder phase modulators and the directional couplers are used to build the readout matrix in the output layer. The performance of the system is studied by a commercial software of optical fiber telecommunication system. The results show that the randomly connected optical neuron networks of the reservoir computing can provide a better performance.
Read full abstract