[Paper Review] Learning to Denoise and Decode: A Novel Residual Neural Network Decoder for Polar Codes

Zhiwei Cao, Hongfei Zhu+2 more|arXiv (Cornell University)|Aug 1, 2019

Error Correcting Code Techniques22 references4 citations

TL;DR

This paper proposes a residual neural network decoder (RNND) for polar codes that integrates a denoising module based on residual learning before neural network decoding. By improving the signal-to-noise ratio of received symbols, the RNND achieves near-optimal bit-error-rate performance with ultra-low latency, outperforming traditional successive cancellation decoding and standard neural network decoders while maintaining comparable computational complexity.

ABSTRACT

Polar codes have been adopted as the control channel coding scheme in the fifth generation new radio (5G NR) standard due to its capacity-achievable property. Traditional polar decoding algorithms such as successive cancellation (SC) suffer from high latency problem because of their sequential decoding nature. Neural network decoder (NND) has been proved to be a candidate for polar decoder since it is capable of oneshot decoding and parallel computing. Whereas, the bit-errorrate (BER) performance of NND is still inferior to that of SC algorithm. In this paper, we propose a residual neural network decoder (RNND) for polar codes. Different from previous works which directly use neural network for decoding symbols received from the channel, the proposed RNND introduces a denoising module based on residual learning before NND. The proposed residual learning denoiser is able to remove remarkable amount of noise from received signals. Numerical results show that our proposed RNND outperforms traditional NND with regard to the BER performance under comparable latency.

Motivation & Objective

• To address the high latency of successive cancellation (SC) decoding in polar codes, which limits real-time wireless communication applications.
• To overcome the suboptimal bit-error-rate (BER) performance of existing neural network decoders (NNDs) despite their low-latency advantage.
• To improve NND performance by introducing a pre-decoding denoising stage using residual learning, enhancing input signal quality before decoding.
• To achieve a balance between BER performance and decoding latency, making the decoder suitable for low-latency 5G NR systems.
• To validate that residual learning-based denoising significantly improves the SNR of received signals, enabling better decoding accuracy.

Proposed method

• The proposed RNND integrates a residual learning-based denoiser that processes noisy received signals before feeding them into a neural network decoder.
• The denoiser is trained end-to-end with the NND using a multi-task learning objective, jointly optimizing for denoising and decoding performance.
• The denoiser uses a residual network architecture to learn the mapping from noisy inputs to cleaner signals, effectively suppressing Gaussian noise.
• The neural network decoder (NND) is implemented using three architectures—MLP, CNN, and RNN—each paired with the denoiser to form RNND variants.
• The entire system is trained using backpropagation with a joint loss function combining denoising and decoding errors.
• The architecture maintains similar parameter counts across NND and RNND variants to ensure fair performance comparison.

Experimental results

Research questions

• RQ1 Can a residual learning-based denoiser improve the BER performance of neural network decoders for polar codes without increasing latency?
• RQ2 How does the pre-decoding denoising stage affect the effective SNR of received signals in polar code decoding?
• RQ3 Does the integration of denoising and decoding in a single RNND framework lead to performance gains over standalone NNDs?
• RQ4 Which neural network architecture (MLP, CNN, RNN) performs best when combined with the proposed denoising module in the RNND?
• RQ5 Can the RNND achieve near-optimal BER performance comparable to successive cancellation (SC) decoding while maintaining ultra-low latency?

Key findings

• The MLP-based RNND achieves a BER gain of approximately 0.2 dB over its NND counterpart at a BER of 1e-4, significantly improving performance.
• The MLP-RNND and RNN-RNND outperform CNN-RNND in both denoising and decoding, with MLP-RNND showing the best overall performance.
• The denoiser improves the effective SNR by about 4 dB at test-SNR = 3 dB, demonstrating substantial noise suppression.
• The MLP-RNND achieves BER performance very close to the successive cancellation (SC) algorithm, which is near-optimal in the given code setting.
• The MLP-RNND runs over 100 times faster than the SC algorithm, making it a strong candidate for low-latency 5G NR applications.
• The computational overhead of the RNND compared to the NND is negligible, with only a slight increase in latency that is practically insignificant.