Pattern Recognition Accuracy Optimization of Unsupervised Spiking Neural Network Using Y-Doped AlN Memristors

Abstract

Inspired by the biological nervous system, the unsupervised spiking neural network (SNN) with the spike-timing-dependent plasticity (STDP) learning rule has been considered as the next-generation artificial neural network. To construct an SNN with high pattern recognition accuracy, hardware with balance synaptic behavior needs to be developed. Here, yttrium (Y)-doped aluminum nitride (AlN) memristors were proposed as artificial synapses in SNNs to investigate the dependence between the doping concentration and the pattern recognition accuracy. With the doping of Y in AlN films, both the memory characteristics and synaptic behaviors of the AlN memristors were optimized. In addition, the STDP parameters of the memristors were extracted and fed into the SNN system to simulate the pattern recognition capability. The optimized pattern recognition accuracies of 75.89% and 60.21% for the MNIST and ETH-80 datasets, respectively, were achieved for the AlN memristors with a Y-doping concentration of 3.4%, which is promising for implementation in future neuromorphic computing system and artificial intelligence.