Scalable Optical Convolutional Neural Networks For Edge Applications

Venkata Anirudh Puligandla, Vladimir Ceperic, Tihomir Knezevic; Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) Workshops, 2025, pp. 1735-1744

Due to the ever-increasing compute and energy requirements of artificial neural networks, the optical neural network (ONN), owing to its high inference speed and low energy requirement, is emerging as an alternative platform for neuromorphic computing. Existing ONN architectures have a high area overhead limiting their applications in edge computing. We propose a simple frequency domain optical convolutional neural network (CNN) architecture using only the Mach-Zehnder interferometer. Our architecture performs similarly to existing architectures using fewer components. The proposed optical convolution layer architecture reduces the area of the photonic integrated circuit by up to 12x. Through ONN simulations, we propose an optical-electric hybrid convolutional neural network architecture to accelerate convolution operations in edge devices that achieves a speedup factor of 2.5x in inference time than a spatial CNN operating on an embedded GPU, and up to 7.8x lower power consumption than a fully-optical CNN architecture.