Photonic extreme learning machine based on phase-sensitive optical-time-domain reflectometry of Rayleigh scattering from a long fiber

Photonic extreme learning machines (ELMs) offer rapid training, low power consumption, and inherent parallelism compared to conventional electronic systems. Time-domain photonic ELMs employing Rayleigh backscattering in an optical fiber for high-dimensional mapping enable simplified architectures and ultrafast processing. In this work, we introduce a novel phase-sensitive optical-time-domain reflectometry method that encodes input data in pulse phase and employs heterodyne detection to eliminate the need for optical amplification of the Rayleigh backscattered signal. In the proof-of-concept experiments, we showcase competitive classification accuracies of 94.93% and 91.56% on the Iris and Vowel datasets, respectively, highlighting the effectiveness and practical advantages of our approach.