Photonic Extreme Learning Machines using Event-Based detection

Abstract

Photonic extreme learning machines use random optical propagation, detection nonlinearity, and a trained linear readout for energy-efficient and scalable optical computing. However, conventional intensity readout with CCD or CMOS cameras constrain the dimensionality of the hidden representation space. Here, we experimentally replace intensity detection with event-based camera, whose thresholded log-intensity response provides alternative pixel-wise hidden representations: first-event time, binary activation, and event count. In nonlinear two spiral classification task we obtain accuracies of $93\pm 3\%$ with strong intrinsic generalization, and comparable ridge and pseudo-inverse performance indicating sample-limited effective dimensionality. Regression results reveal sensitivity to systematic optical-intensity drift, identifying stability requirements for future event-based PELMs. These results establish event-based detection as a route toward richer photonic hidden representations while clarifying current limitations.