Digital-analog hybrid matrix multiplication processor for optical neural networks

Abstract

Optical neural networks (ONNs) promise computing efficiency beyond microelectronics for modern artificial intelligence (AI). Current ONNs using analog matrix-vector multiplication (MVM) implementations are fundamentally limited in numerical precision due to accumulated noise in electro-optical processing. We propose a digital-analog hybrid MVM architecture that achieves a high numerical precision without sacrificing computing efficiency. Our fabricated proof-of-concept hybrid optical processor (HOP) achieves 16-bit precision in high-definition image processing, with a pixel error rate of 1.8 × 10−3 at a signal-to-noise ratio of 18.2 dB, and shows no accuracy loss in MNIST digit recognition. We further explore applying the HOP processor in You Look Only Once (YOLO) object detection and demonstrate sufficient numerical precision is crucial for high confidence detection in real-world neural networks. The hybrid optical computing concept may be applied to various photonic MVM implementations to enable accurate optical computing architectures.