Deep optoelectronic reservoir computing using electrically cascaded semiconductor lasers

Laser dynamics induced by optical perturbations are often unstable when suffering from phase fluctuation or polarization mismatch. Conventional laser-based deep reservoir computers rely on optical injection locking to form dynamically stable steady states for hidden layers, the risks of optical perturbations are inevitable in practice. In this work, a deep optoelectronic reservoir using electrically cascaded semiconductor lasers is proposed and numerically investigated. On the one hand, each layer consists of a laser subject to delayed electrical feedback from the photodetection of its own emission. On the other hand, the feed-forward couplings between layers are achieved through electrical cascading. Therefore, the laser dynamics for each layer inherently avoid the influences from phase fluctuation and polarization mismatch. To verify the effectiveness of deep architecture, the computing performance is evaluated by a task-independent indicator as well as self- and cross-prediction benchmark tasks. The accuracy of predictions is effectively enhanced by increasing the number of layers, independent of the task types. Besides, better performance is observed in the negative feedback regime, which agrees with the enhancement of complexity. The deep optoelectronic reservoir also shows good tolerances of performance improvements against parameter mismatches between hidden layers, which contributes to the flexibility in hardware implementations.