Optical neuromorphic computing based on a large-scale laterally coupled laser array

Abstract

We propose a neuromorphic photonic time-delayed reservoir computing (RC) scheme based on a large-scale laterally coupled laser array with optical feedback and injection. Here, the feasibility of this RC system for dealing with a complex task-image recognition is proved numerically. By the offline pre-processing using randomly initialized convolutional neural network, two common datasets in the visual domain, i.e., MNIST and Fashion-MNIST, are employed to evaluate the performance of the proposed RC. After enlarging the laser array size and optimizing the parameters, our RC system achieves identification accuracy of 98.2% for the MNIST dataset and of 89.9% for the Fashion-MNIST dataset. Since the feedback delay time and information mask period of each laser in the array are coincident and maintained at 0.2 ns, the information processing rate of the RC system reaches 5 Gbps. Besides, the potential situations of parameter mismatch are also analyzed.