Improving the Stability for Spiking Neural Networks Using Anti-noise Learning Rule

摘要

Most of the existing SNNs only consider training the noise-free data. However, noise extensively exists in actual SNNs. The stability of networks is affected by noise perturbation during the training period. Therefore, one research challenge is to improve the stability and produce reliable outputs under the present of noises. In this paper, the training method and the exponential method are employed to enhance the neural network ability of noise tolerance. The comparison of conventional and anti-noise SNNs under various tasks shows that the anti-noise SNN can significantly improve the noise tolerance capability.