Visualizing and Understanding Optoelectronic Neural Networks via the Orbital Angular Momentum of Light

Abstract

As deep learning is penetrating almost every field of research and industry, the demand for computational improvements drives us to investigate new computing paradigms, such as optical and hybrid optoelectronic neural networks. Despite continuous efforts in both free-space and on-chip pathways, scant attention was paid to algorithms suitable for these new hardware. Here we introduce a model visualization method for diffractive optoelectronic neural network (see Fig. 1(a)), proposed in [1]. Specifically, we implement the visualization in a scenario of measuring orbital angular spectrum (OAM) of vortex light, a challenging task in OAM community. Our findings may inspire the link between intelligent computing and physical effects and applications.