An LDPC Decoding Method Using a Neural Network as a De-Mapper

In this paper, we propose a novel LDPC decoding model that employs a fully connected neural network as a demapper to achieve excellent decoding performance while maintaining low computational complexity and computational effort. This neural network model does not need to rely on external coding during the training process, thus significantly reducing the required computational resources. In addition, the model can quickly converge to a parameter configuration with excellent performance through a limited number of iterations. To further enhance the decoding results, this paper also combines this neural demapper with a modified version of the confidence propagation (BP) algorithm. The integrated model achieves excellent decoding performance while keeping the computational load low, even under more stringent conditions.