SNNOT: Spiking Neural Network With On-Chip Training for MIMO-OFDM Symbol Detection

Abstract

Advancing 5G and beyond communications require innovative solutions for symbol detection in MIMO-OFDM. Our research leverages Spiking Neural Networks (SNNs), surpassing the efficiency of traditional Deep Neural Networks (DNNs) and Artificial Neural Networks (ANNs) in dynamic wireless contexts. The SNNOT architecture features a novel triplet Spike-Timing-Dependent Plasticity (STDP) learning circuit, utilizing 22FDX technology from GlobalFoundries, which overcomes STDP’s supervised learning limitations and enhances dynamic learning by converting spikes into voltages. This innovation leads to substantial efficiency gains, with a minimal energy requirement of $20.92 \mu W$ and a small silicon footprint of $37 \times 62 \mu m^{2}$ . Performance evaluations show the triplet STDP rule in SNNOT considerably improves image classification and symbol detection, with over 3% error rate reduction in MNIST and CIFAR-10. Additionally, it achieves a 0.07 symbol error rate (SER) at 5 dB $E_{b}/N_{o}$ in 4 pulse amplitude modulation (4-PAM) Gaussian channel symbol detection, outperforming other methods. With 8-PAM, it surpasses other methods by at least 0.03 SER and maintains lower or comparable SERs in scenarios using the WINNER II channel model and 16 quadrature amplitude modulation (16-QAM). Importantly, our method offers up to 32% faster processing, providing a potent, efficient solution for receiver processing in cutting-edge wireless systems.