Learning to Decode Double Polar Codes for Joint Source-Channel Coding

In this paper, a neural belief propagation decoder of double polar codes for joint source-channel coding (JSCC) is proposed. In the decoder, a residual structure is designed to establish information interaction mechanisms at different sizes between the neural sub-decoders and within them. To compensate for the information loss caused by source compression, a charger layer is introduced. In addition, we designed an iterative training scheme and parameter-sharing mechanism to constrain the complexity of the decoder. The multi-loss is extended to JSCC for training the decoder. It constrains the distribution of latent variables, thereby guiding the ResBP to optimize in the expected direction and accelerating training convergence. Experimental results show that the proposed neural decoder outperforms the existing BP-based decoders of double polar codes.