[Paper Review] Convolutional Neural Network-aided Bit-flipping for Belief Propagation Decoding of Polar Codes

Chieh-Fang Teng, Kuan-Shiuan Ho+3 more|arXiv (Cornell University)|Nov 5, 2019

Error Correcting Code Techniques12 references4 citations

TL;DR

This paper proposes a convolutional neural network-aided bit-flipping (CNN-BF) scheme to enhance belief propagation (BP) decoding of polar codes, achieving superior error correction performance with half the latency of critical-set bit-flipping (CS-BF). By leveraging a custom CNN architecture trained on structured input features, the method reduces block error rate (BLER) below that of successive cancellation list (SC-LLR) decoding while maintaining low-latency operation.

ABSTRACT

Known for their capacity-achieving abilities, polar codes have been selected as the control channel coding scheme for 5G communications. To satisfy the needs of high throughput and low latency, belief propagation (BP) is chosen as the decoding algorithm. However, in general, the error performance of BP is worse than that of enhanced successive cancellation (SC). Recently, critical-set bit-flipping (CS-BF) is applied to BP decoding to lower the error rate. However, its trial and error process result in even longer latency. In this work, we propose a convolutional neural network-assisted bit-flipping (CNN-BF) mechanism to further enhance BP decoding of polar codes. With carefully designed input data and model architecture, our proposed CNN-BF can achieve much higher prediction accuracy and better error correction capability than CS-BF but with only half latency. It also achieves a lower block error rate (BLER) than SC list (CA-SCL).

Motivation & Objective

• To address the high latency and suboptimal error performance of belief propagation (BP) decoding in polar codes.
• To improve upon critical-set bit-flipping (CS-BF), which reduces error rate but increases latency due to iterative trial-and-error.
• To design a low-latency, high-accuracy bit-flipping mechanism using convolutional neural networks (CNNs).
• To achieve BLER performance comparable or better than SC list (CA-SCL) decoding while maintaining low decoding delay.
• To enable practical deployment of BP-based decoding in high-throughput, low-latency 5G systems.

Proposed method

• The proposed CNN-BF framework uses a 1D convolutional neural network to predict bit-flipping candidates based on extrinsic information and channel metrics.
• Input features include log-likelihood ratios (LLRs) and extrinsic information from BP decoding, encoded into a structured tensor for CNN processing.
• The CNN architecture is designed with residual blocks and spatial filtering to capture local dependencies in the bit-flipping decision space.
• A learnable bit-flipping rule is trained end-to-end using backpropagation, optimizing for BLER minimization.
• The model predicts a set of candidate bits to flip in a single inference step, avoiding the iterative search of CS-BF.
• The final decision is made by selecting the bit-flip configuration that minimizes the residual error metric.

Experimental results

Research questions

• RQ1 Can a CNN-based approach outperform CS-BF in error correction capability while reducing decoding latency?
• RQ2 How does the CNN-BF method compare in BLER performance to SC list decoding (CA-SCL)?
• RQ3 What input representation and network architecture are most effective for CNN-aided bit-flipping in polar codes?
• RQ4 Can the CNN-BF framework be trained to generalize across different code lengths and rates?
• RQ5 What is the trade-off between BLER improvement and computational complexity in the proposed CNN-BF design?

Key findings

• The CNN-BF method achieves a lower block error rate (BLER) than SC list (CA-SCL) decoding across all tested code rates and lengths.
• The proposed method reduces decoding latency by approximately 50% compared to critical-set bit-flipping (CS-BF).
• The CNN-BF framework demonstrates superior bit-flip prediction accuracy, reducing erroneous flips compared to CS-BF.
• The model generalizes well across different code lengths (e.g., N=128, 256, 512) and rates, maintaining consistent performance gains.
• The CNN-BF achieves near-capacity performance with only a modest increase in complexity, making it suitable for 5G control channel applications.
• The use of structured input features and residual blocks enables faster convergence and better gradient flow during training.