Unsupervised Learning of Temporal Constancies by Pyramidal-Type Neurons.

Abstract

An unsupervised learning principle is proposed for individual neurons with complex synaptic structure and dynamical input. The learning goal is a neuronal response to temporal constancies: If some input patterns often occur in close temporal succession, then the neuron should respond either to all of them or to none. It is shown that linear threshold neurons can achieve this learning goal, if each synapse stores not only a weight, but also a short-term memory trace. The online learning process requires no biologically implausible interactions. The sequence of temporal associations can be interpreted as a random walk on the state transition graph of the input dynamics. In numerical simulations the learning process turned out to be robust against parameter changes.