On Joint Neural Min-Sum Decoding and Quantization Optimization

Low-density parity-check (LDPC) codes are integral to modern communication systems, including the 5G standard, which supports a new service type known as massive machine type communications (mMTC). mMTC terminals, typically battery-powered and designed for short message transmissions, may face performance challenges with LDPC codes, which excel at longer lengths but show worse performance at shorter ones. Advanced decoders have been developed to address performance issues. One of the approaches enhances the belief propagation (BP)-based decoders by introducing trainable parameters, resulting in the neural BP decoders. To mitigate the energy consumption caused by the decoding complexity, message quantization is also employed in practice. This paper addresses both issues by proposing the collaborative utilization of a neural min-sum decoder to enhance decoding performance and a neural quantizer to reduce computational and memory demands. We present a joint training strategy that achieves superior performance compared to conventional methods.