Deep Learning Based Digital Backpropagation Enabling SNR Gain at Low Complexity

Abstract

A computationally efficient deep learning based digital backpropagation (DL-DBP) algorithm providing a 1.9 dB SNR over a conventional linear compensation (chromatic dispersion compensation algorithm) and a 1 dB gain over a conventional back-propagation algorithm of the same complexity is presented. The algorithm has been tested in a 1200km transmission experiment. Also, if the algorithm is tested against a conventional digital backpropagation algorithm with the gain, then the new algorithm requires a factor 6 lower complexity. We discuss its training procedure and its principle. We discuss its training procedure and its principle.