Optical Neural Network Architecture in Photoferroelectrics

Abstract

We present a novel, versatile optoelectronic neural network architecture for implementing supervised learning in photo-ferroelectrics (Sr x Ba1 − x Nb2O6, Bi12XO20; X=Ge, Si, Ti, LiNbO3: Fe, LiTaO3:Fe and LiTaO3:Cr). The system is based on spatial multiplexing rather than the more commonly used angular multiplexing of interconnect gratings. This sample, single-crystal architecture implements a variety of multi-propagations and Marr-Albus-Kanerva style algorithms. Extensive simulations show how suitable modified supervised learning algorithms compensate for beam depletion, re-scattering, absorption and decay effects of the crystals. This type of implementation also benefits strongly from recently discovered tunable stability phenomena in the read-write properties of anisotropic crystals like photo-ferroelectric.