A Deep Learning based Decoder for Polar Codes with Large Kernels

This dissertation implements a deep learning based SC-decoder for polar codes with kernel of order 16, experimenting with a flexible design that in principle can be applied to any polar code, with any block-length or code-rate and with a kernel of any size. This provides a practical and flexible method to attack the open problem of decoding large-kernel polar codes in real-time. Although the decoder sacrifices some performance (allowing for a slightly increased probability of errors in the decoded messages), it provides a significant reduction in the decoding time. To some extent, the trade-off between performance and complexity can be calibrated by changing the size and the sparsity of the neural networks in the decoder. To showcase the flexibility of the method, the dissertation applies it also to the problem of decoding the extended binary Golay code. Preliminarily, this requires to choose a suitable polarizing kernel of order 24 and to freeze the first 12 channels, so that the resulting polar code coincides with the Golay [24,12,8]-code.