Backpropagation for Continuous Theta Neuron Networks
Abstract
The Theta neuron model is a spiking neuron model which, unlike traditional Leaky-Integrate-and-Fire neurons, can model spike latencies, threshold adaptation, bistability of resting and tonic firing states, and more. Previous work on learning rules for networks of theta neurons includes the derivation of a spike-timing based backpropagation algorithm for multilayer feedforward networks. However, this learning rule is only applicable to a fixed number of spikes per neuron, and is unable to take into account the effects of synaptic dynamics. In this thesis a novel backpropagation learning rule for theta neuron networks is derived which incorporates synaptic dynamics, is applicable to changing numbers of spikes per neuron, and does not explicitly depend on spike-timing. The learning rule is successfully applied to XOR, cosine and sinc function mappings, and comparisons between other learning rules for spiking neural networks are made. The algorithm achieves 97.8 percent training performance and 96.7 percent test performance on the Fischer-Iris dataset, which is comparable to other spiking neural network learning rules. The algorithm also achieves 99.0 percent training performance and 99.14 percent test performance on the Wisconsin Breast Cancer dataset, which is better than the compared spiking neural network learning rules.
Citation
Fan, David Dawei (2015). Backpropagation for Continuous Theta Neuron Networks. Master's thesis, Texas A&M University. Available electronically from https : / /hdl .handle .net /1969 .1 /186998.