A Physical Neural Network Based On Coupled In-plane Laser Diodes

Abstract

Semiconductor lasers can be used as physical reservoir computers (RCs) that can perform ML tasks faster and more efficiently than traditional neural networks. However, most setups rely on fiber-optic delay lines or external cavities which hinder scalability. This thesis uses numerical simulation to investigate laser-based RC setups which do not rely on time delays to remember past inputs, instead exploiting lasers’ relaxation oscillations. An analysis of a a reservoir computer composed of uncoupled lasers demonstrated strong linear memory and state-of-the-art performance on the NARMA-10 benchmark. To increase nonlinear processing capacity, unidirectionally coupled laser chain reservoirs were evaluated, showing improved accuracy but requiring an impractically large number of lasers. A novel binary tree laser reservoir topology was developed, where pairs of dissimilar lasers couple into subsequent layers to exponentially increase nonlinear interactions. This binary tree architecture achieved state-of-the-art performance on the NARMA-10 benchmark (NMSE = 0.100) utilizing a total of only 63 lasers, requiring just 32 lasers in the first layer to be directly driven by the modulation current.