Unsupervised Learning Feature Estimation for MISO Beamforming by Using Spiking Neural Networks

Abstract

Deep learning (DL)-based beamforming frameworks have the advantage of low computational complexity. However, existing works are based on artificial neural networks (ANNs) and have high energy cost. To improve this situation, this letter proposes a feature estimation framework based on spiking neural networks (SNNs) for sum rate maximization (SRM). Specifically, we first devise a coding method at the input side to convert channel state information into binary spike sequences that can be processed by SNNs. Then, at the output side, the output binary sequences are decoded into the features that are used to generate the beamforming weights. Based on the characteristics of the decoding method, unsupervised learning can be used to train the SNNs. Simulation results show that compared to existing feature estimation frameworks using ANNs, the SNNs-based framework reduces energy cost by orders of magnitude and has comparable sum rate performance.