Photoelectric hybrid convolution neural network with coherent nanophotonic circuits

To achieve low-power convolutional neural networks, we develop a photoelectric hybrid neural network (PHNN), which consists of the optical interference unit (OIU) and field-programmable gate array (FPGA). The OIU composed of Mach–Zehnder interferometers (MZI) arrays, used as convolution kernels, performs multiplication and accumulation operations. The convolution kernel is split and reorganized, forming a new unitary matrix, which reduces MZI quantity. FPGA realizes nonlinear calculation, data scheduling and storage, and phase encoding and modulation. Our PHNN has an accuracy rate of 88.79%, and the energy efficiency ratio is 1.73 times that of traditional electronic products.