A modified supervised learning rule for training a photonic spiking neural network to recognize digital patterns

Abstract

A modified supervised learning rule which is suitable for training photonic spiking neural networks (SNN) is proposed for the first time. The proposed learning rule is independent of the time intervals between actual spike and desired spike or between presynaptic spike and postsynaptic spike. Based on the proposed supervised learning rule, 10 digital images are learned in photonic neural network which consists of 30 presynaptic neurons and 10 postsynaptic neurons. Presynaptic and postsynaptic neurons are photonic neurons based on vertical-cavity surface-emitting lasers with an embedded saturable absorber (VCSEL-SA). The results show that 10 digital images are recognized correctly in photonic SNN after enough training. Additionally, the effects of learning rate, the jitters of learning rate, initial weights distribution of SNN and bias current of postsynaptic neurons (VCSELs-SA) on the recognized error are examined carefully based on the proposed learning rule. To the best of our knowledge, such modified supervised learning rule has not yet been reported, which would contribute to training photonic neural networks, and hence is interesting for neuromorphic photonic systems and pattern recognition.