Illustration of Convallis rule as applied to speech data.
(A) Preprocessing pipeline. Waveforms corresponding to utterances of eleven spoken digits (zero to nine plus “oh”) by multiple speakers were processed by a cochleogram model , which was used to produce inhomogeneous Poisson spike trains of 100 input cells. (B) Illustration of spiking and voltage responses after learning for two particular digits. Top row: examples of the input population spike patterns corresponding to a single presentation of the digits “four” and “five”. Top row right, pseudocolor representation of the simulated neuron's input weights after learning, for rate constraint, the Convallis rule, and rcSTDP. Bottom three rows show a raster representation of the trained neuron's responses to a test set consisting of 300 utterances of these digits by previously unheard speakers, together with a membrane potential trace from a single test-set utterance. Right column shows mean firing rate vs. time averaged over the whole test set, illustrating the development of selective responses by the Convallis rule.