Feedforward and Recurrent Neural Network-Based Transfer Learning for Nonlinear Equalization in Short-Reach Optical Links

Abstract

Neural network (NN)-based nonlinear equalizers have been shown effective for various types of short-reach direct detection systems. However, they work best for a certain channel condition and need to be trained again when the channel environment is changed, which hinders the efficient deployment of future optical switched data center networks. In this article, we propose transfer learning (TL)-aided feedforward neural networks (FNN) and recurrent neural networks (RNN) for nonlinear equalization in short-reach direct detection optical links, which enables a fast transition to new equalizers when the channel condition is changed. A 50-Gb/s 20-km pulse amplitude modulation (PAM)-4 optical link is experimentally demonstrated as the target system, and links of varying bit-rates and fiber lengths are selected as the source system. Experimental results show that TL could help reduce the number of epochs and training symbols of FNNs/RNNs required for nonlinear equalization in the target system, taking advantage of FNNs/RNNs trained for source systems. A reduction of 90%/87.5% in epochs and 62.5%/53.8% in training symbols is achieved with FNNs/RNNs transferred from the most similar source system. We also find that FNNs can be transferred to their corresponding RNNs for equalization in the target system, while TL from RNNs to FNNs cannot work properly. TL enables a fast transition between different NN-based equalizers, which is critical for future optical switched data center networks, where the optical links need to be dynamically reconfigured.