Deep Time-Delay Reservoir Computing With Cascading Injection-Locked Lasers

Abstract

Reservoir computing is a simplified recurrent neural network, which requires training only at the output layer and hence significantly reduces the training cost. Time-delay reservoir computing (TDRC) introduces a large number of virtual neurons based on a single physical neuron and a feedback loop, which is friendly for hardware implementations. This work proposes a scheme for implementing the deep TDRC architecture based on cascading injection-locked semiconductor lasers. In each layer, the reservoir consists of a quantum dot laser and an optical feedback loop. The output of each reservoir layer is fed into the subsequent one through the optical injection-locking technique. This all-optical approach (for reservoir layers) has the merit of high scalability without any depth limitation. Theoretical analysis shows that the deep TDRC improves the performance on multiple benchmark tasks, including the memory capacity, the prediction of chaos, the nonlinear channel equalization, and the recognition of spoken digits.