Here is a short excerpt of a lecture on Autoencoder Communication Systems and the general topic of “Learning to communicate”. The lecture was recorded on May 19th, 2022 and is part of the course “Deep Learning Applications in Communications” at the Institute of Telecommunications, University of Stuttgart.

In the video you can see me explaining the training process of a bit-wise autoencoder communications system. The system tries to find optimal signals to transmit k=4 bit within n=2 real valued channel uses at the training SNR of 3.85dB. On the left side one can see the currently used constellation and on the right side one can see the mutual information obtained at receiver side. The bottom heat map plots show the current decision regions of the learned receiver, where pink regions indicate a strong decision towards bit 0 and blue regions indicate a strong decision towards bit 1. It can be seen, that the system finally converges towards a constellation that differs from conventional modulation schemes, like the comparable 16-QAM constellation, and is thereby able to transmit more information at the operational point of 3.85dB. While these geometrical shaping gains are well known in communications research, the possible game changer of a neural network based autoencoder system is, that it is able to optimize its signals for any channel environment, including all unknown insufficiencies, while solely relying on data based training strategies.

For more information see the paper “Trainable Communication Systems: Concepts and Prototype” by S. Cammerer, F. A. Aoudia, S. Dörner, M. Stark, J. Hoydis and S. ten Brink at ieeexplore.ieee.org where these results are based on.