Paper on bit error and block error rate training for ML-assisted communication now available online!

Next generation communications will rely significantly on machine learning (ML). But how can we get the most out of ML-assisted communication systems?
In this new work, the authors show that popular loss functions such as binary cross entropy (BCE) or mean squared error (MSE) are well-suited for the training of bit error rate (BER) optimal communication systems. However, the authors also show that the BCE and MSE loss are intrinsically unable to learn the conditional dependencies between multiple bits, which are required for block error rate (BLER) optimal decoding. Since BLER (and not BER) is the relevant figure of merit in virtually all practical communication applications, the authors propose new loss functions that are aimed at minimizing the BLER.

The experiments show that the difference in BLER performance between training with BCE or MSE (solid curves) and training with a BLER-specific loss (dashed) can be as large as 0.78 dB, even when all other parameters (such as the channel, the transceivers, and the number of training samples) are fixed.

The paper was co-​authored with Chris Dick (NVIDIA), Haochuan Song (Southeast University), and Prof. Christoph Studer (IIP Group). A preprint of this work is available on external page arXiv.