Neural network concatenation for Polar Codes

TL;DR

This paper introduces an efficient algorithm for neural network decoding of polar codes of any size, achieving near-ML performance with manageable training complexity, overcoming previous size limitations.

Contribution

It presents a novel algorithm enabling neural network decoding for large polar codes with performance comparable to ML decoding, reducing training complexity.

Findings

Decoding performance is comparable or better than SC decoding.

The algorithm scales efficiently to large polar codes.

Neural network decoding approaches can be practical for longer codes.

Abstract

When a neural network (NN) is used to decode a polar code, its training complexity scales exponentially as the code block size (or to be precise, as a number of message bits) increases. Therefore, existing solutions that use a neural network for polar decoders are stuck with short block sizes like 16 or 32. Despite the fact that the NN training is very complex for long polar codes, the NN decoding gives the better latency and its performance is potentially close to the maximum likelihood (ML). In this paper, we describe an efficient algorithm to create the NN decoding for a polar code of any size with the initial performance that is equal or better than that of successive cancelation (SC). Therefore, it creates an opportunity to design the NN based decoding with the performance that is as close to the ML, as the training time allows.