SIGNETS: Neural Network Architectures for m-QAM Soft Demodulation

This paper presents a novel approach to Quadrature Amplitude Modulation (QAM) demodulation using neural networks, addressing the limitations of traditional demodulation techniques in complex channel conditions. Through systematic Neural Architecture Search and Hyper-parameter optimization, we develop a family of Convolutional Neural Network architectures that demonstrate robust performance across challenging channel conditions, including multi-path fading, inter-symbol interference, and non-linear distortions, without requiring explicit channel estimation. We comparatively evaluate the members of the family of networks to find a Pareto-optimal Neural Network Demodulator with a balance of demodulation accuracy and computational cost, achieving an average accuracy of 99.658% across 4 dB to 24 dB SNR while requiring less than 16,000 Floating-Point Operations (FLOPs) for every demodulated QAM-16 symbol. A practical Field Programmable Gate Array (FPGA) implementation that achieves 2.52 Million bits per second throughput while maintaining 99.55% demodulation accuracy through structured pruning and quantization-aware training is presented. Experimental validation over acoustic channels demonstrates superior performance compared to traditional techniques, with performance further enhanced through fine-tuning.