It's not a major part of the paper, but Google tested a neural network decoder (which had the highest accuracy), and some of their other decoders used priors that were found using reinforcement learning (again for greater accuracy).