Fully differential decoder for decoding lattice codes using neural networks

Abstract

Short-length lattice codes are crucial in various applications, including channel estimation and quantization. This paper introduces a novel weighted lattice decoder (WLD) that utilizes a parametric function to process decoder inputs and incorporates a weighted Belief Propagation (BP) algorithm. To further enhance the accuracy of the decoder's estimations, a new two-part multiloss function is proposed. This innovative approach significantly improves the performance of $E_8$, Barns-Wall ${\text{BW}}_8$, and BCH lattice codes. The proposed WLD demonstrates notable improvements in the error-floor region, achieving gains of up to 1.4 dB and 2.3 dB on the Symbol Error Rate (SER) curve compared to the primary BP decoder and the Neural Network Lattice Decoding Algorithm, respectively. By leveraging these advancements, the WLD offers a more robust and efficient decoding solution, making it highly suitable for real-time applications where low latency and high accuracy are paramount.