Dataset Conversion for Spiking Neural Networks

Spiking Neural Networks (SNN) are still a relatively new research area, and there are only a few publicly available datasets suitable for SNN training and testing. Datasets designed for traditional neural networks are not suitable as an input to SNNs because they rely on the timing of spikes, thus, requiring data to be preprocessed in a specific way. This work presents a pipeline for converting existing conventional datasets into rate-encoded spikes matching the requirements for SNN processing. The implementation is based on the Python snntorch library. The functionality of the proposed encoder pipeline is evaluated on a music genre classification task using the GTZAN dataset. However, the proposed procedure has general applicability and may be used for data of any form.