Photorefractive Integrated Photonics for Analog Signal Processing in AI

Abstract

The computational cost of AI could be alleviated by accelerating the synaptic transfer calculations in artificial neural networks with an analog crossbar array processor. In this work, we present the core building blocks of an all-optical integrated photorefractive crossbar array for artificial neural network training by demonstrating photorefractive synapses in an integrated 2-D beam interaction network. We show that the photorefractive quality of the circuits resembles that of the bulk GaAs crystal that they were fabricated from. Then, this work experimentally validates the integrated photorefractive crossbar array design and constitutes a framework for engineering photorefractive integrated photonics.