STDP enables spiking neurons to detect hidden causes of their inputs

Description

Institute for Theoretical Computer Science, Graz University of TechnologyThe principles by which spiking neurons contribute to the astounding computationalpower of generic cortical microcircuits, and how spike-timing-dependentplasticity (STDP) of synaptic weights could generate and maintain this computationalfunction, are unknown. We show here that STDP, in conjunction witha stochastic soft winner-take-all (WTA) circuit, induces spiking neurons to generatethrough their synaptic weights implicit internal models for subclasses (orcauses) of the high-dimensional spike patterns of hundreds of pre-synaptic neurons.Hence these neurons will fire after learning whenever the current input bestmatches their internal model. The resulting computational function of soft WTAcircuits, a common network motif of cortical microcircuits, could therefore bea drastic dimensionality reduction of information streams, together with the autonomouscreation of internal models for the probability distributions of their inputpatterns. We show that the autonomous generation and maintenance of thiscomputational function can be explained on the basis of rigorous mathematicalprinciples. In particular, we show that STDP is able to approximate a stochasticonline Expectation-Maximization (EM) algorithm for modeling the input data. Acorresponding result is shown for Hebbian learning in artificial neural networks.