Improving detection in optical communications using all-optical reservoir computing

During the last decade reservoir computing (RC) has proven to be a powerful machine learning concept using recurrent networks to process sequential information streams [1]. Lately, the concept of RC was drastically simplified to reduce hardware requirements, thus making machine learning concepts more obtainable to apply in actual physical systems [2]. RC has been implemented in photonic configurations and has efficientìy processed analog optical signals following simple schemes and minimal training overhead. Especially the realization using a semiconductor laser with time-delayed optical feedback has shown the potential of all-optical machine learning, utilizing signals with transient states at the GHz regime [3]. In this work we exploit the properties of a small-scale RC in order to improve the detection capabilities of lightwave communication signals of very low quality. The RC has a limited number of virtual nodes (N=33) and a total internal time-delay of 1.65ns in order to allow fast post-processing. Its design follows an all-optical configuration, including a discrete-mode semiconductor laser (SL) that emits at 1550nm and an optical feedback cavity, as presented in [3]. We numerically simulate the lightwave communication signals by considering an onoff keying modulated binary stream at 10Gb/s that has undergone 2000 km of fiber transmission (20 segments of 100-km spools including optical amplification and dispersion compensation). At such distances, direct threshold detection results in BER values higher than 0.1. Even when applying offline post-processing with a ridge regression training algorithm, which considers the patterns of the detected bits within the complete detection period, the detection BER value improves only slightly to 0.031. When we employ the RC, the transmitted signal is optically injected into the RC's SL, resulting in a nonlinearly transformed output which depends directly on the operating conditions of the reservoir. After training the reservoir with the previously considered regression algorithm, we find many operating conditions for which the bit stream recovery potential is improved significantly (Fig. 1a). Especially when the memory properties of the reservoir are incorporated in the training process - i.e. by taking into account the transformed patterns of the previous bits - BER values lower than 10-4 have been retrieved (Fig. 1b).