Band-Structure-Engineered Electronic-Photonic Nonlinear Activation Functions
Zheheng Xu, David P. Burghoff · Physical Review Applied · 2022
Fast, sensitive, and compact devices that implement nonlinear activation functions are needed to form fully connected photonic neural networks (PNNs). However, even in highly nonlinear media, optical nonlinearities are relatively weak. We propose here a scheme for implementing nonlinear activation functions that relies on band-structure-engineered nanostructures. This scheme realizes the smallest possible hybrid optoelectronic approach, relying on fast electronic processes to implement nonlinearity instead of a true optical nonlinearity. Using well-established simplified density-matrix models, we demonstrate architectures that exhibit a low-intensity threshold of $3.5\phantom{\rule{0.2em}{0ex}}\ensuremath{\mu}\mathrm{W}$ along with a fast optical response of 10 ps in a relatively small linear footprint of $4\phantom{\rule{0.2em}{0ex}}\ensuremath{\mu}\mathrm{m}$. We also show that PNN training performance is improved in handwritten pattern recognition when applying our simulated nonlinear activation function, indicating potential for creating deep fully connected PNNs.