Robustness of classification ability of spiking neural networks

The robustness of an artificial neural network is important for its application. In this paper, we focus on the robustness of the classification ability of spiking neural networks with respect to perturbation of inputs according to the probability distribution. Two typical types of perturbations, sinusoidal and Gaussian perturbations, are considered, which have rarely been investigated for SNNs in the existing literature. In particular, some of the perturbations are allowed to be large, rather than all the perturbations are uniformly small as in the existing literature. Numerical experiments are carried out by using the SpikeProp algorithm on classical XOR problem and other three benchmark datasets. The numerical results show that the classification ability of SNN is robust with respect to sinusoidal and Gaussian perturbations of input signals.