Problem

Recurrent neural networks (RNNs) are highly effective for temporal tasks and video processing, but their implementation on conventional electronic hardware consumes significant energy due to repetitive memory operations. This overhead is a critical constraint for applications requiring efficient temporal processing, such as edge computing, autonomous systems, and real-time sensor analysis, where power budgets and processing latency are stringent. While photonic computing offers advantages in parallelism and low-latency, extending it to recurrent temporal processing, which necessitates memory and feedback, often reintroduces electronic control, negating the benefits photonics aims to provide.

Method

This research introduces a novel method for implementing recurrent neural networks through the passive spatiotemporal propagation of light in multimode optical fibers. Video frames are encoded as phase masks onto spatially separated regions of a spatial light modulator (SLM). These encoded beams are directed into multimode fibers of varying lengths to introduce controlled time delays, ensuring temporally ordered inputs arrive sequentially. The combined optical field then enters a 50/50 fiber coupler, with half of the signal recirculating through a fiber loop to establish a fading memory. The other half is directed to a camera for readout as high-dimensional speckle patterns. The system relies on intrinsic spatiotemporal wave dynamics, including interference and Kerr-type nonlinear light-matter interactions within the multimode fiber, to generate complex, high-dimensional states that encode both current inputs and past information. Critically, the entire optical setup remains fixed, with no trainable parameters or electronic feedback required for the recurrent dynamics. Learning is performed solely by a linear electronic readout layer that processes the optical output states.

Results

The optical recurrent processor demonstrated competitive performance across a variety of temporal and spatiotemporal learning tasks. For chaotic time-series forecasting using the Santa Fe benchmark, the system achieved a mean squared error (MSE) of 0.084 on the test set, capable of accurate long-horizon predictions. In human action recognition on the KTH dataset, it reached 98.33% classification accuracy, comparable to established video recognition methods, and showed robust encoding of action, scene, and identity. For autonomous driving, the system successfully extracted temporal context from road scenes, producing smooth steering angle predictions that closely tracked ground-truth signals. Finally, in surgical skill assessment using the JIGSAWS dataset, the optical reservoir accurately classified overall skill levels and specific performance metrics. These results highlight that a single, fixed physical optical configuration can achieve competitive performance across diverse tasks without modifying the optical architecture.

Implications

This work demonstrates a significant paradigm shift, showing that recurrent computation can emerge directly from the spatiotemporal physics of light propagation in multimode fibers, rather than being algorithmically implemented. This physical recurrence offers an energy-efficient pathway to temporal artificial intelligence by leveraging intrinsic optical nonlinearities and passive propagation. The system generates high-dimensional optical states that inherently encode current inputs and a fading memory without requiring digital memory or trained recurrent parameters. The passive nature of the recurrent loop enhances stability by reducing sensitivity to timing jitter and thermal drift, and the energy cost per recurrent update is minimal. While the current implementation requires precise optical alignment and has a fixed memory horizon, the findings point towards a broader class of learning systems where physical dynamics handle the core recurrent computation, with electronic processing limited to readout and training. Future advancements in photonic integration and programmable optical elements could lead to more compact and scalable physical recurrent processors.