A Noncoherent Incremental Learning Demodulator

Abstract

Incremental learning after deployment is one of several attractive capabilities that motivate the use of neural network demodulators. This paper presents a complex noncoherent neural network suitable for on-off key (OOK) demodulation. When trained in an AWGN channel, the demodulator learns a solution that outperforms the traditional noncoherent matched filter demodulator. The paper also explores incremental learning techniques that enable continued learning in the field. Training in the field with known labels provides maximum adaptability to new conditions, but the availability of known symbols maybe limited. As an alternative, we considered the effectiveness of entropy regularization and pseudo-labels to adapt a lab-trained reference network to new field conditions. Simulation of these techniques in an example multipath channel demonstrates successful unsupervised adaptation with initial symbol error rates up to 20% and successful semi-supervised adaptation with a small fraction of known symbols per packet and initial symbol error rates as high as 40%. In both cases, symbol error rates after adaptation are below 0.3%.