An efficient discrete model for implementing temporal coding spiking neural network

Spiking neurons, which makes up SNNs are more biologically realistic artificial neurons which convey information by the timing of spikes. SNNs have been applied for various learning tasks and they are capable to perform with improved accuracy in less number of training cycles. Due to the temporal nature a spiking neuron cannot be implemented as a standard neuron. Generally spiking neurons are implemented as a discrete model and the output is computed through an iterative process. Here, each iteration is assumed to take one unit time. Even though the learning models of SNNs can learn in less number of cycles, they consume more computational resources due to the large number of iterations they require. This paper proposes a numerical approach using Newton-Raphson method to estimate the firing time of a spiking neuron in lesser number of iterations. Proposed model was implemented and analysed to test its accuracy, convergence and capability for learning. Tests were performed using the Iris plants dataset and Breast cancer dataset from UCI Machine learning repository. The proposed model was able to find the actual neuron firing time within only a few iterations. Tests have shown that for the proposed approach the required iterations needed to estimate the firing times to an accuracy of two decimal places would be at least 400 fold lesser compared to the standard approach. Number of required iterations found to be independent of the firing time in contrast to the standard approach. The model also found to be smoothly converges in finding the actual firing time. Further the model was tested to cluster Iris and Cancer data sets and the accuracy was found to be 92.7% and 97.4 % respectively. The results demonstrate that the limitation of estimating the actual firing time with less computational resources can be resolved using the proposed discrete approach. This would pave way to devise better learning algorithms for SNNs with even higher accuracy and low training cycles.