Publication:
Reconfigurable Deep Photonic Recurrent Networks for Optoelectronic Information Processing
Date
2026
Proceedings Paper
Loading...
Journal
MACHINE LEARNING IN PHOTONICS II
Abstract
We demonstrate reconfigurable deep photonic recurrent networks using an 8-input/8-output cascaded Mach-Zehnder interferometer mesh that generates wavelength-dependent nonlinear projections, and an optoelectronic feedback loop that re-injects photo-detected signals to achieve fading-memory dynamics. Evaluated on the MIT-BIH Arrhythmia database across 14 rhythm classes, single-shot and recurrent modes achieve 94.38% and 97.51% classification accuracy respectively, representing an improvement of up to 7.5-percentage-points over the digital baseline. For chaotic sequence forecasting and nonlinear regression, the network reaches R2 = 0.9934 (MSE = 0.0006) and R2 = 0.9988 (MSE = 0.0004), respectively. We show that the network performance improves monotonically with recurrence depth across all tasks, confirming that the optoelectronic feedback mechanism contributes systematic representational capacity beyond what is achievable by forward-pass-only systems. This architecture offers a compact, wavelength-scalable, and reconfigurable platform for optical information processing built on the deep photonic network design platform.