Digital electronic neural networks with analog nanophotonic frontends: A numerical study

There is recent interest in integrating optics into neural computing systems considering the potential energy savings and speed enhancements. Optical neural networks relying on free-space diffractive effects can potentially permit the use of low resolution imagers and off load a portion of electronic computation. They are of interest in microscopy and edge computing applications. A numerical analysis of ideal few-layered diffractive optical neural networks with trainable phase masks (in real and Fourier spaces) under monochromatic coherent illumination is presented. Stacked all-dielectric transmissive metasurfaces are particularly suitable for the realization of such networks at optical frequencies. Six different kinds of networks (with and without Fourier space phase masks; electronic dense layers; and intensity-dependent optical nonlinearity) are considered for a comparative analysis. The networks are assessed on their testing accuracy on the MNIST, Fashion MNIST and the grayscale CiFAR-10 datasets. Optical variants were found to perform well on MNIST (≥96% testing accuracy) and F-MNIST (≥85%) datasets where objects are clearly demarcated from the background. However, the optical variants performed poorly on the CiFAR dataset (≈44%) in comparison to state-of-the-art electronic deep convolutional neural networks. The presence of electronic dense layers or optical nonlinearity provided marginal improvements of about 2% in test accuracy. Furthermore, network scaling by widening the phase plates and/or cascading more layers is found to only marginally improve test accuracy.