Optimizing DMD embeddings for optical computing with physicsinspired data preprocessing

Abstract

Free-space optical computing platforms employing spatial light modulators are constrained by speed owing to the low refresh rate. As a solution, relatively faster digital micromirror devices are employed to increase the data throughput, but they cause information loss because of their 1-bit nature. We overcome the information loss caused by digital micromirror devices in optical computing platforms by optimizing a mapping based on random projection and thresholding algorithms. The physics-inspired framework preprocesses input samples prior to deployment on photonic hardware accelerators for high-TOPS inference.