Temporal Pattern Classification using Spiking Neural Networks
Abstract
A novel supervised learning-rule is derived for Spiking Neural Networks (SNNs) using the gradient descent method, which can be applied on networks with a multi-layered architecture. All existing learning-rules for SNNs limit the spiking neurons to fire only once. Our algorithm however is specially designed to cope with neurons that fire multiple spikes, taking full advantage of the capabilities of spiking neurons. SNNs are well-suited for the processing of temporal data, because of their dynamic nature, and with our learning rule they can now be used for classification tasks on temporal patterns. We show this by successfully applying the algorithm on a task of lipreading, which involves the classification of video-fragments of spoken words. We also show that the computational power of a one-layered SNN is even greater than was assumed, by showing that it can compute the Exclusive-OR function, as opposed to conventional neural networks.
keywords: spiking neural networks, temporal pattern recognition, classification, gradient descent.
Downloads
gzipped postscript of thesis (614 kb), pdf of thesis (719 kb).
The slides (152 kb) I used at my defense can also be downloaded.
Thanks to Sander Bohte I also presented my work at the CWI . Here are the slides (159 kb) I used.
