Sparse Neural Network for Detection and Decoding of Non-Binary Polar-Coded SCMA

Sparse code multiple access (SCMA) and polar codes have demonstrated superiority in supporting massive connections and short-packets, which are essential in the uplink transmission of the internet of things. This paper proposes a sparse neural network (SNN) for the detection and decoding of the non-binary polar-coded SCMA. The network consists of the message passing algorithms (MPA) and the sparse belief propagation (SBP) modules, which are immigrated from the MPA and BP algorithms, respectively. The nodes of the bipartite BP factor graph that do not contribute to the belief propagation are pruned to form the SBP module. A structure factor
$\alpha $
defined by the ratio of the number of layers in the MPA modules and that of the SBP modules is analyzed to find a balance between the decoding performance and the complexity. The weights are set on the neurons of the layers and trained by the stochastic gradient descent. Finally, the simulation results show substantial improvements in bit error rate with low complexity over the traditional receiver. Specifically, the SNN receiver outperforms the traditional receiver by 2.56 dB in the AWGN channel when the code length is 16, the code rate is
$\frac {1}{2}$
, and the user overloading factor is 150%.