End-to-End Learning System for Symbol Decoding in Optical Communication

In this paper we propose an autoencoder-based End-to-End (E2E) Deep Learning for symbol decoding in fiber optical communications systems. The autoencoder was trained with data encoded in two different ways. First, one-hot encoding which corresponds to the most common approach to train the network in order to optimize the decision regions of each symbol. Then, we introduce the binary coding where the bit sequences that represent the symbols, are positioned in such a way to minimize the error and compensate the distortions introduced by the optical channel. The focus of the study are the generated constellations, symbol decision regions and symbol bit labeling. The autoencoder contributes to better visualization and analysis of the resulting decision regions for scenarios with high and low nonlinearities. It improves the interpretability of the symbol decoding. To the best of our knowledge, this is the first time when a ML-based approach (i.e the autoencoder) estimates not only the ideal locations of the decoded symbols in various constellations (4, 16, 64) but also optimizes the geometrical decision regions of the symbols in the constellation.