A Graph-Neural-Network Decoder with MLP-based Processing Cells for Polar Codes

Compared with traditional decoders, the Neural Network Decoder (NND) has the advantages of being low-latency, high-throughput and one-shot. However, the conventional deep neural network decoder (DNND) usually suffers the computational mismatch and the lack of generalization capability. In this paper, we propose a novel Graph-Neural-Network Decoder for the popular Polar Codes, which is directly constructed based on the regular decoding graph of Polar Codes and by replacing each of its basic 2-by-2 polarization elements with a simple Multi-Layer-Perceptron (MLP) based processing cell. The resultant decoder, namely PC-GNND, is thus endowed with the ability to infer over the skeleton of the decoding graph and has a greatly improved generalization capability. Simulation results show that, the proposed PC-GNND is capable of learning the exact code structure as well as the channel noise very efficiently with only a tiny fraction of the entire codebook and achieving better performance than that of conventional NNDs with far less parameters and training epochs. Moreover, the PC-GNND trained in a particular block length can be scaled to another one for different block lengths after fine tuning, which significantly reduces the computational cost of training.