DIR-Net: Deep Residual Polar Decoding Network Based on Information Refinement.

Area of Science:

• Information Theory
• Coding Theory
• Deep Learning

Background:

• Polar codes approach the Shannon limit but traditional decoding methods face latency and throughput issues.
• Deep learning offers potential for improved polar code decoding, but current black-box approaches struggle with distinguishing valid from interfering information, limiting Bit Error Rate (BER) performance.

Purpose of the Study:

• To propose a novel deep residual network, DIR-Net, for decoding polar-coded short packets.
• To enhance the accuracy and reduce the bit error rate of polar code decoding by effectively distinguishing between effective and interference information.

Main Methods:

• A two-stage decoding network comprising a denoising subnetwork and a decoding subnetwork.
• Construction of the entire network using an attention mechanism for enhanced information extraction.
• Cascaded attention modules for step-by-step information filtering and refinement.

Main Results:

• DIR-Net effectively distinguishes between effective and interference information within codewords.
• The proposed two-stage network structure improves decoding accuracy.
• DIR-Net demonstrates superior Bit Error Rate (BER) performance compared to existing decoding methods.

Conclusions:

• DIR-Net offers a significant advancement in polar code decoding for short packets.
• The attention-based, information-refining approach leads to lower bit error rates.
• The method shows robust performance across both Additive White Gaussian Noise (AWGN) and flat fading channels.