Hybrid training of optical neural networks

Optical neural networks are emerging as a promising type of machine learning hardware capable energy-efficient, parallel computation. Today’s optical mainly developed to perform inference after in silico training on digital simulators. However, various physical imperfections that cannot be accurately modeled may lead the notorious “reality gap” between simulator and system. To address this challenge, we demonstrate hybrid where weight matrix is trained with neuron activation functions computed optically via forward propagation through network. We examine efficacy three different networks: an linear classifier, opto-electronic network, complex-valued study comparative training, our results show robust against kinds static noise. Our platform-agnostic scheme can applied wide variety networks, work paves way towards advanced all-optical intelligence.