A Hybrid Photonic–Electronic Inception CNNs with Nonlinear APMR Convolutions⋆

Abstract

Abstract: This paper introduces a novel, hybrid Inception CNN that uses all-pass microring resonators (APMRs) to perform optical convolution and nonlinear activation. The APMR-Inception
module combines an optical branch with electronic branches to keep the model compact while still
capturing multi-scale features. Experiments show stable training and good accuracy: around 85% on
CIFAR-10, 95.2% on SVHN, and 62% on CIFAR-100. The model works well on structured objects
and remains balanced performance across classes. Because many operations are done through passive optical interference, the system could reduce energy use and latency. Overall, the results indicate
that nonlinear APMR-based photonic modules are a promising approach for building efficient deep
learning models.
Keywords: All-pass microring resonator (APMR), Photonic neural networks, Hybrid
optical–electronic CNN, Energy-efficient optical computing.