Converting Data for Spiking Neural Network Training

Abstract

The application of spiking neural networks (SNNs) for processing visual and auditory data necessitate the conversion of traditional neural network datasets into a format suitable for spike-based computations. Existing datasets designed for conventional neural networks are incompatible with SNNs due to their reliance on spike timing and specific preprocessing requirements. This paper introduces a comprehensive pipeline that enables the conversion of common datasets into rate-coded spikes, meeting processing demands of SNNs. The proposed solution is evaluated on Spike-CNN trained on Time-to-First-Spike encoded MNIST and compared with the similar system trained on the neuromorphic dataset (N-MNIST). Both systems have comparative precision; however the proposed solution is more energy efficient than the system based on neuromorphic computing. Since, the proposed solution is not limited to any specific data form and can be applied to various types of audio/visual content. By providing a means to adapt existing datasets, this research facilitates the exploration and advancement of SNNs across different domains.