Abstract
Precise neuronal spike timing plays an important role in many aspects of cognitive processing. Here, we explore how a spiking neural network can learn to generate temporally precise spikes in response to a spatio-temporal pattern, through spike-timing-dependent plasticity modulated by a delayed reward signal. An escape noise neuron is implemented as the readout to incorporate the effect of background noise on spike timing. We compare the performance of two different escape rate functions that drive spiking in the readout neuron: the Arrhenius & Current (A&C) and Exponential (EXP) model. Our results show that the network can learn to reproduce target spike patterns containing between 1 and 10 spikes with 10 ms temporal accuracy. We also demonstrate the superior performance of the A&C model over the EXP model for the parameters we consider, especially when reproducing a large number of target spikes.