Photonic dimensionality reduction front-end for integrated optical neural networks

Abstract

We present a photonic dimensionality-reduction front-end for silicon photonic integrated optical neural networks (ONNs) that addresses the mismatch between full-resolution images, which contain hundreds to thousands of pixels, and current integrated ONNs, which typically support only tens of input channels. The front-end combines a Fourier lens with an on-chip two-dimensional grating-coupler array: the lens maps the input image to the spatial-frequency domain, and the grating couplers coherently sample the resulting Fourier-plane field. This enables ONNs to process optically compressed input directly, without electronic preprocessing or on-chip optical-signal generation. We experimentally implement the front-end and acquire dimensionality-reduced representations from MNIST and grayscale CIFAR-10. When these measured representations are used as inputs to representative simulated ONN classifiers, they retain near-baseline performance for binary tasks, achieving F1 scores of at least ≥ 90 % for MNIST and ≥ 74 % for CIFAR-10 while using only 2–5 % and 2–6 % of the original image pixels, respectively. This establishes optical Fourier domain dimensionality reduction as a practical front-end strategy for bridging full-resolution visual inputs with limited input dimensionality of integrated ONNs.