Propagation characteristics of weak signal in feedforward Izhikevich neural networks

Abstract

The feedforward neural network is widely applied in various machine learning architectures, in which the synaptic weight between layers plays an important role in the weak signal propagation. In this paper, the five-layer Izhikevich neural networks with excitatory or excitatory–inhibition neurons are employed to study the effect of Gaussian white noise and synaptic weight between layers on the weak signal transmission characteristics of the subthreshold excitatory postsynaptic currents signal imposed on the input layer. It can be found that there is an optimum value of noise intensity (a medium noise intensity) at which the weak signal can be stably propagated in the five-layer Izhikevich neural networks. The spike timing precision (STP) under the optimal noise intensity will become maximum by increasing the synaptic weight. The noise intensity and synaptic weight corresponding to the maximum value of the STP in the excitatory–inhibition network are smaller than those in excitatory–inhibition network. For the smaller or the bigger noise intensity, however, the STP will become very small, and the weak signal cannot be transmitted from the input layer to the output layer. Furthermore, the weak signal is propagated from the input layer to the output layer and enhanced under a larger synaptic weight in the feedforward neural networks.