Low Complexity Convolutional Neural Networks for Equalization in Optical Fiber Transmission

Mitigation of these distortions is possible using, e.g., digital back-propagation (DBP) [1]. DBP, however, requires knowledge of fiber parameters and topology, and can be computationally expensive in part due to potentially large number of spatial segments. Artificial neural networks (ANNs) offer an alternative approach that might be less complex [2]. Learned digital backpropagation (LDBP) is proposed in [3], in which the model is based on the split-step Fourier method (SSFM), and optimized using the standard learning algorithms for ANNs. In this paper, we consider a convolutional neural network (CNN) equalizer. We note that the SSFM coefficients are repeated in each span. Thus, instead of joint training of all neural network layers, we train a few unique layers that are shared in network depth. With this parameter sharing method, we substantially reduce the number of model trainable parameters.