Recurrent Network with Attention for Symbol Detection in Communication Systems

Abstract

One major challenge for wireless receivers to maintain information fidelity involves the demodulation of faded signals in noisy environments. Typical demodulation techniques for M-ary quadrature amplitude modulated (M-QAM) signal utilize variants of coherent demodulation. This paper explores deep learning (DL), specifically by using a proposed architecture recurrent-attention networks to compliment, or even overcome the limitations of demodulating M-QAM symbols. The proposed model is shown to outperform the benchmark coherent demodulator and other DL-based demodulators such as convolutional neural network (CNN), recurrent neural network (RNN) and the hybrid of both up to 5 dB learning gain at a lower model complexity and requires less memory usage.