A Low-complexity Neural BP Decoder with Network Pruning

Existing deep learning-based channel decoders, called neural decoders, suffer from demands on an excessively high computational complexity and large memory resource. In this work, we will show that a low-complexity neural belief propagation (BP) decoder can be constructed by utilizing the network pruning technique. In particular, it will be shown that by removing unimportant edges in a neural BP decoder, a significant complexity gain can be achieved. When the decoding complexity is fixed, the proposed decoder highly achieves a notable performance improvement as compared to the existing neural BP decoder, which will be demonstrated with performance evaluations. In addition, we conduct a preliminary study investigating the structure of pruned edges, which we believe provides some clues of a general design framework of practical neural BP decoders.