Neural Network Decoding of Polar Codes with Large Kernels

Polar codes with large kernels offer improved error-correction performance but suffer from high decoding complexity due to costly marginalization operations. In this work, we propose neural network (NN)-based approximations of these marginalization functions, enabling efficient decoding for arbitrary linear kernels. We show that training on all-zero partial codewords is sufficient for linear kernels, significantly reducing the dataset size and allowing for smaller, faster NN models. Our method is evaluated within a Successive Cancellation (SC) decoder, being compatible with any blocklength, code rate, or SNR. Experimental results for kernel size 16 demonstrate strong performance and substantial complexity savings over exact marginalization.