Time-Delayed Reservoir Computing System Based on a Three-Stage Monolithic Integrated Amplified Feedback Laser With Electrical Information Injection

Reservoir computing (RC) system based on integrated devices is highly charming in large neural network hardware systems. In this work, we propose and experimentally investigate a time-delayed RC (TD-RC) system based on a three-stage monolithic integrated amplified feedback laser (AFL) with electrical information injection. The AFL in this TD-RC system is composed of a 220 mu m long distributed feedback (DFB) laser section, a 240 mu m long passive phase section, a 320 mu m long amplifier section and two 20 mu m long isolators for electronic isolation among above three sections, and then the total feedback length is 600 mu m. The information to be processed is sent into the reservoir layer by phase-modulating the passive phase section, and the temporal outputs of the AFL are repeatedly extracted and taken as the virtual node states of the TD-RC for training and testing. Via a Santa-Fe time series prediction task, the effects of the node interval and the sampling rate on the prediction performance are explored, and the results show that the normalized mean square error (NMSE) can reach 0.0569 at a processing speed of 0.1 GSa/s.