A Hypothesis on How the Neocortex Extracts Information for Prediction in Sequence Learning

Abstract

From the biological view, each component of a temporal sequence is represented by neural code in cortical areas of different orders. In whatever order areas, minicolumns divide a component into sub-components and parallel process them. Thus a minicolumn is a functional unit. Its layer IV neurons form a network where cell assemblies for sub-components form. Then layer III neurons are triggered and feed back to layer IV. Considering the delay, through Hebbian learning the connections from layer III to layer IV can associate a sub-component to the next. One sub-component may link multiple following sub-components plus itself, but the prediction is deterministic by a mechanism involving competition and threshold dynamic. So instead of learning the whole sequence, minicolumns selectively extract information. Information for complex concepts are distributed in multiple minicolumns, and long time thinking are in the form of integrated dynamics in the whole cortex, including recurrent activity.