Spiking neural networks and their applications

Artificial neural networks (ANNs) have been developed as adaptable, robust function approximators for at least the last quarter-century. They have progressed through two generations, and the third is now under development. Spiking neural networks (SNNs) seek to improve on previous generations in two ways: by using a more biologically-inspired neuron, they are shown to be capable of more complex calculations; incorporating polychronous properties of highly-recurrent networks with delays of different lengths on each synapse to achieve large numbers of possible patterns with relatively few neurons and synapses. Abstracted spiking neurons have been used as a third-generation activation function in a traditional feedforward network architecture, and their potency in application to a real-world problem -- identification of power system generator dynamics -- is demonstrated in this dissertation in comparison to a standard sigmoidal multi-layer perceptron network. However, the goal of SNNs is to be able to utilize biological-like neural network modeling to capture the computational prowess of living brains. In order to achieve such a feat, first a bio-inspired SNN must be able to handle continuous-valued function approximation; until this is done, such networks cannot even be compared to their second-generation predecessors. This dissertation demonstrates a technique for using a faithfully modeled SNN on continuous-valued problems. The encoding and decoding frameworks developed in this dissertation for the biologically-inspired SNN enables it, like any other ANN, to be applied to any time-dependent problem, including neuroidentification of power system generator dynamics --Abstract, page iii