Optical Speckle Patterns as Nonlinear Feature Maps for Photonic Machine Learning

Abstract

The rapid growth of artificial intelligence (AI) has increased the demand for energy-efficient hardware, and optical computing has emerged as a promising alternative to conventional electronic processors owing to its high parallelism and potentially lower power consumption. In this work, we investigate an optical computing architecture based on an integrated photonic circular multipass cell, focusing on its suitability for machine learning and pattern recognition tasks. The properties of the system are evaluated through measures of nonlinearity and cosine similarity, which characterize the system's information-processing capacity and the dimensionality of the generated feature representations. Numerical results demonstrate that the reservoir mirror reflectance, the size of the encoded figure, and the refractive-index contrast between the reservoir and the encoded figure modify the nonlinearity of the system, as verified via pairwise similarity estimates. These results indicate that the features generated by the circular multipass cell can be tailored by adjusting the system parameters, providing a means to control the reservoir transformation of the input data. The resulting representations were applied to classification tasks involving texture images and thyroid collagen images, achieving classification accuracies ranging from 69% to 93%. These findings suggest that photonic computing systems could provide a basis for more energy-efficient machine-learning hardware.