Individual-Driven Spiking-Mixer Deep Learning Model for IRS-Assisted mmWave Systems Beam Training

Abstract

The impacts of noise and channel variation are critical challenges for beam training in millimeter wave (mmWave) intelligent reflecting surface (IRS) systems. This paper proposes a novel hybrid deep learning (DL) scheme containing individual spiking encoders and a feature mixer network named the individual-driven spiking-mixer (IDSM) DL model. The proposed spiking encoder converts channel intensities into binary sequences by the leaky integrate-and-fire (LIF) mechanism for a robust channel representation. Moreover, the feature mixer network obtains the optimal beam through two views of firing tokens quantified from binary sequences. Experimental results show that our proposal can achieve reliable beam prediction and a high channel capacity with low complexity.