Passive photonic recurrent neural networks via spatiotemporal multimode fiber dynamics

Abstract

Reservoir computing enables efficient processing of temporal data by exploiting the intrinsic dynamics of physical systems while requiring training only at a linear readout layer. Photonic implementations are particularly attractive due to their inherent parallelism, high bandwidth, and low latency. Here, we experimentally demonstrate a passive recurrent photonic reservoir based on multimode fiber dynamics for video recognition and chaotic time-series forecasting. Input data are encoded onto a phase-only spatial light modulator and propagated through a multimode fiber, where modal dispersion and interference generate high-dimensional reservoir states. Temporal recurrence is introduced via optical reinjection, providing memory without active nonlinear components. The resulting speckle patterns are recorded by a camera and processed using a trained linear readout. The system is evaluated on multiple benchmarks, including human action recognition, surgical skill classification, steering-angle prediction, and closed-loop forecasting of the Santa Fe laser time series.