Abstract
Artificial neural networks (ANNs) have greatly advanced the field of video and image processing. These machine learning algorithms have been proven to excel at tasks such as object detection and handwritten digit recognition. This is shown in Wan et al.’s (2013) approach that performs handwritten digit recognition on the MNIST data set with 0.21% error [1]. Although ANNs have achieved incredibly high levels of accuracy on these tasks when simulated in traditional computing environments, there is an increasing demand for performing these tasks in real time on embedded computer systems with low power consumption.
Researchers believe that spiking neural networks (SNNs) are a suitable alternative because of their energy efficiency and event-driven architectures [3]. However, more research is required on SNNs to determine the best neural models, encoding methods and training techniques for their use in applications such as image processing applications.
Our group expands on the research of Deng et al. (2020) in comparing ANNs to SNNs for image classification tasks [4]. Deng et al. (2020) compare ANNs to SNNs by comparing their accuracy in classifying samples from the MNIST benchmark. Our group reimplements these tests in the BindsNET framework [5] using different neural models, encoding methods, and training techniques to study how these factors affect the SNN model accuracy.