Iterative Syndrome-Based Deep Neural Network Decoding

Abstract

While the application of deep neural networks (DNNs) for channel decoding is a well-researched topic, most studies focus on hard output decoding, potentially restricting the practical application of such decoders in real communication systems. Modern receivers require iterative decoders, a pivotal criterion for which is the ability to produce soft output. In this paper, we focus on this property. We begin by modifying the syndrome-based DNN-decoding approach proposed by Bennatan et al. (2018). The DNN model is trained to provide soft output and replicate the maximum a posteriori probability decoder. To assess the quality of the proposed decoder’s soft output, we examine the iterative decoding method, specifically the turbo product code (TPC) with extended BCH (eBCH) codes as its component codes. A sequential training procedure for optimizing the behavior of component decoders is utilized. We illustrate that the described approach achieves exceptional performance results and is applicable for iterative codes with larger code lengths $[n=4096, k=2025]$ , compared to state-of-the-art DNN-based methods. Finally, we address the issues of computational complexity and memory requirements of DNN-based decoding by analyzing the model’s compression limits through pruning and matrix decomposition methods.

Keywords

• Channel decoding
• complexity reduction
• deep neural networks
• iterative codes
• soft-output