Optical Convolutional Neural Network with Atomic Non-linearity

Convolutional neural networks (CNNs) have become an indispensable part of machine learning due to their successful application on image-based machine learning problems. However, the compute resources required to train and infer such CNNs scale quadratically with problem size, thus expense of time and energy is large. Unlike electrons, photons can be used to naturally realize massive and parallel interconnections [1] , achieve clock speeds in the GHz range and most importantly the computational cost for optical convolutions scales only linear with problem size. Thus, optical computing is getting ever more attractive [2] .