Deep-Learning based Equalization of Highly Compressed Faster Than Nyquist Signals

In this paper, we investigate the decoding of Fasterthan-Nyquist (FTN) signals using Recurrent Neural Networks (RNNs) and compare it to traditional decision feedback equalization (DFE). We demonstrate the near-optimal performance of the RNN equalizers and their ability to decode FTN signals at significant compression factors where traditional DFE fails. The use of these RNN equalizers allows us to reliably communicate up to twice as many bits as the Nyquist signalling system at the same power and with nearly the same error probabilities at low complexity.