Comparison of Output Decoding Techniques for Spiking Neural Networks

Abstract: Spiking Neural Networks (SNNs) hold significant potential due to their high energy efficiency when implemented on specialized hardware. Central to SNNs is the translation of sequences of spike events to concrete outputs, like class predictions. This study conducts a comparative analysis of two popular output decoding methods: rate coding, where the class corresponding to the output neuron with the most spikes is selected, and temporal coding, which predicts the class of the neuron that spikes first. SNNs with a common architecture, using these decoding methods, were tested on two datasets. The results reveal that temporal coding exhibited higher variance across experiments, was more affected by the sample size, and had considerably lower accuracy on the smaller dataset compared to rate coding. This difference may stem from the high sensitivity of temporal coding to early spikes, altering network predictions. However, further examination with diverse SNN architectures and datasets is required to validate the results from this report.