Design and Analysis of Neural-Network-based, Single-User Codes for Multiuser Channels

Inspired by its success in other fields, there have been many recent developments in the use of machine learning and neural networks to enable multiuser communication and to design efficient channel codes along with practical decoders. However, there has been little attempt to combine the results of these efforts. In this paper, for the first time, we present a neural network autoencoder architecture to jointly address both problems. The resulting codes designed by our simple and easy-to-train neural network can have arbitrary rates, are comparable to existing state-of-the-art neural network designed codes, and are directly applicable in a multiuser context. We analyze these single-user codes and characterize the design parameters which affect their performance. We then show that these same single-user codes can be used to operate close the maximum sum rate of a K-user Gaussian multiple access channel (MAC) under various SNR scenarios, without the need for retraining or learning a joint code. This improved performance is achieved by introducing a new iterative successive interference cancellation method (SIC) that outperforms traditional onion-peeling.