Learning method for neural networks using weight perturbation of orthogonal bit sequence and its application to adaptive WDM demultiplexer

This paper proposes a novel on-chip learning method for hardware-implemented neural networks to achieve an adaptive wavelength division multiplexer (WDM) demultiplexer. The parameters of the neural network are perturbed by orthogonal bit sequences with small amplitude. The parameters are corrected based on the correlation detection result between the perturbed error signal and the corresponding perturbation signal. A learning experiment that transmits 200-Mb/s, four-channel WDM signals through a 40-km fiber and the tracking of the wavelength drift of the optical transmitter successfully demonstrate the proposed method