Multi-channel free-space optical convolutions with incoherent light

Abstract

Free-space optical systems are promising candidates for high performance computing and have been particularly successful in the implementation of large-scale convolutions. Convolutions are the key operation in convolutional layers, which are used extensively in modern neural networks, especially in the context of image/video processing and generation. These optical accelerators have demonstrated remarkable performance in both processing rates and energy efficiency. Prior approaches have primarily demonstrated convolutions from a single input channel to one or more output channels. We extend these methods to perform true multi-channel convolutions, where multiple input channels are convolved with their own sets of convolutional kernels onto output channels. We simulate this approach using both ray-tracing and angular spectrum propagation and find the approach is highly-scalable. We then experimentally implement a proof-of-concept prototype to demonstrate multi-channel free-space optical convolutions.