Simulation of In-Situ Training in Spike Neural Network Based on Non-Ideal Memristors

In this paper, we proposed a multilayer supervised training method based on Multi-Resume and memristor synaptic properties. We use a new differential equation describe the dynamic of Pt/HfO2/Al2O3/Ti devices and update weights based on present conductance. Specifically, the weight update is achieved by applying just one pulse to the device, which will simplify the peripheral circuits. In addition, the algorithm training method, asymmetric nonlinear weight update and synaptic variation measured from experiments are investigated for the impact on network accuracy. Our results show that the nonlinearity of devices does not much affect the network accuracy; The hybrid training is a better method for ensuring the accuracy; The spiking neural network (SNN) shows remarkable high tolerance to the variation of the device. This work will lay the foundation for later on-chip learning of SNN based on memristors. (Optdigits dataset: https://archive.ics.uci.edu/ml)