Training Process of Memristor-Based Spiking Neural Networks For Non-linearity

The implementation of a spiking neural network (SNN) with memristor arrays has the potential to improve the area and power efficiency of edge-device inference. However, due to non-linearity, we cannot maintain the inference accuracy on a memristor array with the trained weights based on a regular SNN. In this paper, we proposed a training process to consider the non-linear circuit effects. With the proposed method, the experimental results showed that the accuracy of the trained SNN was improved from 62.99% to 92.81 %.