Hardware-Efficient Optical Matrix Processor via Low-Rank Approximation

We propose a hardware-efficient optical matrix processor based on low-rank approximation, utilizing narrowband filters of microring resonators (MRRs) and broadband Mach-Zehnder interferometers (MZIs). Our design enables the realization of a 4×4 matrix using only four MRRs and four MZIs, achieving a 50% reduction in optical components compared to conventional architectures that rely on crossbar arrays of MRRs or MZI arrays implemented through singular value decomposition. We experimentally validate the efficacy of this approach in image convolution using functional equivalents of the optical matrix processor. Additionally, in a digital number classification task, our proposal achieves a prediction accuracy of 96%, closely matching the theoretical accuracy of 96.21%. These results highlight the efficiency and accuracy of our optical matrix processor design, offering a viable solution for executing large-scale matrix operations with improved resource utilization.