Design considerations for optical deep learning networks

Performing deep learning in the optical domain is attractive due to the very low electrical power requirements when compared to running networks on a GPU. Since a single positive lens can perform a Fourier transform, correlation operations are relatively simple to implement and they have the potential of a very large computational bandwidth. However, many of the current designs of deep learning networks are not easily implemented in the optical domain. In this paper we develop a python framework for simulating optical deep learning using Pytorch. This allows the discovery of the optimal weights by calculating them on a realistic optical system. Noise sources, speckle models, and calibration errors can be accounted for. The effect of readily realisable filters such as nematic liquid crystal phase only spatial light modulators is investigated. Key differences still exist such as activation functions and the ability to modulate the signal is limited.