Efficient Active Deep Decoding of Linear Codes Using Importance Sampling

Abstract

The quality and quantity of training data significantly affect deep learning model performance. In error correction, generating high-quality samples with minimal noise is crucial. This letter presents a method that combines a modified Importance Sampling (IS) distribution with active learning to generate high-quality samples. The suggested IS distribution generates samples iteratively from shells with error probabilities within a specific range. This approach enhances the performance of BCH(63,36) and BCH(63,45) codes with cycle-reduced parity-check matrices. The proposed IS-based-active Weight Belief Propagation (WBP) decoder improves the error-floor region by up to 1.9dB on the BER curve compared to the conventional WBP decoder.