Abstract: A more biologically plausible spiking response model (SRM) is presented to cope with the learning problem of spiking neural networks (SNN) in which neurons can spike multiple times. In constructing this model, the dependence of the postsynaptic potential upon the firing times of the postsynaptic neuron is not neglected. We derive an additional error back-propagation learning rule for the coefficient of the refractoriness function. The algorithm has been tested on classification tasks or XOR problem, IRIS dataset and Poisson spike trains. The results show that the SRM based SNN with neurons that fire multiple spikes can transfer information more efficiently and speed up training compared to SNN with neurons that fire only once.

Key words: Spiking neural networks, multiple spikes, spike response model, refractoriness