Classification Performances due to Asymmetric Nonlinear Weight Updates in Analog Artificial Synapse-Based Hardware Neural Networks

Artificial synapses are fundamental for neuromorphic computing to overcome the bottleneck of the von Neumann system. In particular, a memristor synapse-based neuromorphic system has been known as an optimal device for effectively implementing a hardware neural network. Here, we propose the memristor synapse which shows potentiation and depression process like biological brain mechanisms and investigate the effects of varying the device parameters of nonlinearity and asymmetry on the classification accuracy. We find that the virtual devices with a nonlinearity of less than 10 can be obtained the classification accuracy up to 80%. Our approach demonstrates a practical neuromorphic system based on virtual device on simulation and measured device on experiment and verifies the feasibility of the hardware neural networks.