Neural decoders can outperform classical algorithms for error correction by learning to scale messages in belief propagation networks, suggesting that hybrid approaches combining domain structure with learned parameters are effective for communication systems.
This paper proposes GBPNet, a neural network decoder for error-correcting codes that improves upon classical belief propagation algorithms. By combining learned message scaling with the traditional Tanner graph structure used in coding theory, the decoder achieves better error correction performance while maintaining interpretability through its message-passing design.