Realization of fault tolerance for spiking neural networks with particle swarm optimization

Abstract

The spiking neural network (SNN) model has been an important topic in the past two decades. Many training algorithms, such as SpikeProp, were designed and applied to various applications. However, the fault tolerant ability in SNNs was not fully understood. Based on our study, the SNN model with the classical training objective function cannot even handle the single fault situation, in which one of the hidden neurons is damage. To improve the fault tolerant ability, we design an objective function and utilize the particle swarm optimization approach to minimize it. Simulation results show that our approach is much better than the classical objective function.