Optimized Layered FAIDs of 5G LDPC Codes With RQNN

In this paper, a recurrent quantized neural network with a layered structure is designed to optimize the look-up table (LUT) in the layered finite alphabet iterative decoders (FAIDs). By using straight-through estimators to soften the backward gradient propagation of some low-precision activation functions, the quantization thresholds and quantization levels in the LUTs are optimized as parameters in the network. Performance evaluations on decoding codewords encoded with quasi-cyclic low-density parity-check codes show significant advantages compared to the conventional optimized FAIDs using a generalized flooding structure network. The proposed approach achieves notable performance improvements without increasing the complexity of the decoding algorithm. We also classify the check nodes by row weights of the parity-check matrix, and train different LUTs for different categories of nodes for each iteration. This approach achieves further performance improvements in online decoding and significant complexity reduction compared to the exhaustive search method.