Evaluation of the Accuracy of Converting a Traditional Neural Network into a Pulsed One with Different Methods for Selecting the Neuron Threshold

This work compares two of the most commonly used methods for setting the thresholds of spiking neurons when transferring weights to them from a traditional neural network. Using the example of the spoken digit classification tasks Heidelberg Digits (HD) and Russian Speech Commands (RuSC), it is demonstrated that setting the thresholds by minimizing the differences between the outputs of a traditional neural network layer and the estimated number of spikes in a spiking network allows for higher classification accuracy. However, on the more complex RuSC dataset, the losses incurred when converting a traditional network to a spiking one are greater due to the issues of premature spiking and late input.