Time-delayed reservoir computing using mutually coupled multimode semiconductor laser for high-speed image recognition

Abstract

Reservoir computing (RC), especially time-delayed RC, which derives from recurrent neural network-based models, has the advantages of being easy to implement at the physical level and having a low training cost. Nowadays, time-delayed RC is being used to handle more complex tasks such as image processing. However, for time-delayed RC, the more complex the tasks requires more virtual nodes, resulting in longer delay lines being employed. This leads to a reduction in the RC information processing rate. To overcome this drawback, multimode semiconductor lasers with multiple modes offer a solution. In our work, we use the mutually coupled multimode semiconductor lasers as the physical nodes to construct a time-delayed RC system. Finally, two MC-MSLs were used, each with four modes. It greatly increased the number of virtual nodes at the same information processing rate. By training the RC with input extracted representative features, we have successfully realized parallel processing of the image recognition task and achieved 99.20% and 86.10% accuracies on MNIST and Fashion-MNIST datasets. Given the expansion of multimode semiconductor lasers in longitudinal mode, MC-MSLs RC is expected to enable high-speed processing of more complex tasks.