Research work · 2023

SwiftSpike: An Efficient Software Framework for the Development of Spiking Neural Networks

Genevieve Claire Fahey, Samuel J. Ippolito, Glenn I. Matthews

IEEE International Conference on Omni-layer Intelligent Systems (COINS)

Abstract

Spiking Neural Networks (SNNs) are Machine Learning (ML) algorithms that use sparse, binary, event-driven spikes to propagate information through the network. Coupled with physical neuromorphic processors, SNNs are more energy efficient compared to matrix-based Artificial Neural Network (ANN) solutions, making them well-suited to resource-constrained Internet of Things (IoT) applications that have strict power and local processing requirements. SNN algorithms still require further research to improve their accuracy and allow them to compete with ANNs but their development is hampered by a lack of fast, modular software and simulation frameworks. In this work we present SwiftSpike, an efficient, customisable SNN development framework written in C++ that supports user-defined neuron and synapse models. We validate SwiftSpike against the widely used Brian 2 framework and demonstrate speed-ups of 17.3x on an unsupervised image recognition task trained with Spike-Timing-Dependent Plasticity (STDP).

DOI: 10.1109/coins57856.2023.10189197