Topology Modification and Training for Spiking Neural Networks

Spiking Neural Networks are a promising field with potential for performing machine learning tasks using hardware with very low SWaP costs. These systems have an inherently temporal nature using event-based information transfer with a dynamic state, making Spiking Neural Networks well suited for dynamic, temporal tasks. However, training a spiking neural network to take advantage of the capabilities of such a system is challenging. Existing Evolutionary Optimization algorithms work well to solve small problems very efficiently but struggle to learn larger problems. This work begins by introducing a network compilation process to assemble these small networks into a more complex whole to potentially operate for more complex problems. Further, this work introduces a means of reducing the size of assembled networks through a process called time multiplexing. Finally a modified version of the EONS algorithm is introduced to operate on problems with larger input spaces by creating more densely-connected networks.