Learned codes for broadcast channels with feedback
Yingyao Zhou, Natasha Devroye
TL;DR
The paper addresses finite-blocklength coding for the two-user Gaussian broadcast channel with feedback by proposing GBCF-Lightcode-sep, a symbol-by-symbol learned code that uses time-division in the first two rounds and neural FE/MLP modules to optimize transmission and decoding. By training with a negative log-likelihood loss and a balance regularizer, the scheme achieves superior BLER performance at short blocklengths ($N\le 10$) and demonstrates robustness to noisy feedback, outperforming OL/EOL and TD-Lightcode in key regimes. An initial interpretation reveals the encoder output in round 3 is approximately linearly related to past feedback, offering insight into how the learned code exploits feedback for refinement. Overall, the work extends learned, multi-user feedback codes to the GBCF, providing a practical approach for reliable short-blocklength communications with feedback.
Abstract
We focus on designing error-correcting codes for the symmetric Gaussian broadcast channel with feedback. Feedback not only expands the capacity region of the broadcast channel but also enhances transmission reliability. In this work, we study the construction of learned finite blocklength codes for broadcast channels with feedback. Learned error-correcting codes, in which both the encoder and decoder are jointly trained, have shown impressive performance in point-to-point channels, particularly with noisy feedback. However, few learned schemes exist for multi-user channels. Here, we develop a lightweight code for the broadcast channel with feedback that performs well and operates effectively at short blocklengths.