Spiking neural network unsupervised learning algorithm SCoBUL and its application to extracting informative features from DVS camera signal

A principally new type of video cameras, the so-called DVS (dynamic vision sensors), became commercially available recently. They are capable of enhancing dramatically speed and energy consumption parameters of video signal processing procedures because they send asynchronous stream of spikes indicating brightness change of individual pixels instead of scanned raster data of the whole camera view field. However, this new kind of output signal requires novel algorithms for its processing. Such an algorithm called SCoBUL (spike correlation based unsupervised learning) is described in the present work. SCoBUL uses a one-layer spiking neural network with lateral inhibition for extracting informative features from spike trains in unsupervised learning regime. SCoBUL's most important feature is a generalization of STDP (spike timing dependent plasticity) rule optimized especially for this task.