Spiking Neural Networks with Nonidealities from Memristive Silicon Oxide Devices

Recent years have seen a rapid surge in the application of artificial neural networks in diverse cognitive settings. The augmented computational demands of these structures have led to an interest in new technologies and paradigms. Of all the artificial neural networks, the spiking neural network (SNN) is notable for its capability to imitate the energy-efficient signalling system in the brain. The memristor presents a promising potential for the integration of SNN into hardware, despite certain non-ideal device properties posing a challenge to its implementation. This study involves the simulation of a SNN model utilizing experimental data on silicon oxide. Particularly, it examines the impact of a non-linear weight update on SNN performance. SNNs were shown to possess tolerance for device non-linearity, while the network can simultaneously maintain a high degree of accuracy. These results provide valuable prior information for future implementation of silicon oxide device-based neuromorphic hardware.