Improving DeepPolar Neural Codes via Reliability-Weighted FER-Centric Loss Functions

Channel coding represents a promising application area for neural network (NN)-based techniques. However, because the existing theoretical encoding and decoding algorithms are already highly optimized, developing NN-based methods that surpass traditional designs remains a difficult task. To address the challenge, this paper studies the recently proposed NN-based channel coding framework known as DeepPolar, with the goal of enhancing its frame error rate (FER) performance beyond that of the conventional successive cancellation (SC) decoder. Toward this goal, we introduce reliability-deweighted (RDW) top-kmax loss, RDWp-norm loss, and RDW focal loss functions to prioritize critical bit positions during an extended training curriculum specifically designed to target FER rather than bit error rate (BER). Numerical results indicate that judicious design of these loss functions leads to a significant FER improvement of approximately 0.9 to 1 dB over the original DeepPolar code and 0.3 to 0.4 dB over the traditional polar code with SC decoding at FER level of 10–4, depending on encoding configurations, without sacrificing BER performances. Furthermore, the proposed code designs exhibit performance close to the normal approximation of the finite blocklength capacity, operating merely 1.7 dB away. This demonstrates their considerable potential to advance NN-based polar codes.