[Paper Review] Low-complexity Recurrent Neural Network-based Polar Decoder with Weight Quantization Mechanism

Chieh-Fang Teng, Chen-Hsi Derek Wu+2 more|arXiv (Cornell University)|Oct 29, 2018

Error Correcting Code Techniques13 references4 citations

TL;DR

This paper proposes a low-complexity recurrent neural network (RNN)-based polar decoder with a codebook-based weight quantization mechanism to reduce memory and computational overhead in polar code decoding. By quantizing network weights using a learned codebook, the method achieves 98% memory reduction with only a minor performance degradation, enabling efficient deployment of deep learning decoders in 5G NR systems.

ABSTRACT

Polar codes have drawn much attention and been adopted in 5G New Radio (NR) due to their capacity-achieving performance. Recently, as the emerging deep learning (DL) technique has breakthrough achievements in many fields, neural network decoder was proposed to obtain faster convergence and better performance than belief propagation (BP) decoding. However, neural networks are memory-intensive and hinder the deployment of DL in communication systems. In this work, a low-complexity recurrent neural network (RNN) polar decoder with codebook-based weight quantization is proposed. Our test results show that we can effectively reduce the memory overhead by 98% and alleviate computational complexity with slight performance loss.

Motivation & Objective

• To address the high memory and computational complexity of deep learning-based polar decoders in practical communication systems.
• To enable efficient deployment of neural network decoders in 5G New Radio (NR) by reducing model size and complexity.
• To explore weight quantization techniques that preserve decoding performance while minimizing memory overhead.
• To develop a codebook-based quantization mechanism tailored for recurrent neural network decoders in polar coding.

Proposed method

• A recurrent neural network (RNN) is designed to decode polar codes, leveraging sequential processing for improved convergence.
• A codebook-based weight quantization mechanism is introduced to reduce the precision of network weights, minimizing memory usage.
• The codebook is learned during training and used to map full-precision weights to low-bit representations.
• The quantized RNN decoder is trained end-to-end with a loss function optimized for block error rate minimization.
• The method maintains the RNN's recurrent structure to exploit temporal dependencies in decoding operations.
• Memory and computational complexity are evaluated by comparing full-precision and quantized models under identical decoding performance targets.

Experimental results

Research questions

• RQ1 Can a low-complexity RNN-based decoder achieve near-capacity performance for polar codes with reduced memory footprint?
• RQ2 How effective is codebook-based weight quantization in minimizing memory overhead without significant performance loss?
• RQ3 What is the trade-off between decoding complexity, memory usage, and block error rate in quantized RNN decoders?
• RQ4 Can the proposed quantization mechanism be effectively integrated into recurrent architectures for polar decoding?

Key findings

• The proposed RNN-based decoder with weight quantization reduces memory overhead by 98% compared to full-precision models.
• The performance loss due to quantization is minimal, with only a slight degradation in block error rate.
• The method maintains fast convergence and competitive decoding performance, approaching belief propagation performance.
• The codebook-based quantization enables efficient deployment of deep learning decoders in resource-constrained 5G NR systems.