Artificial Dendrites: an Algorithm

Abstract

This paper outlines preliminary work on an algorithm to create dendritic, tree-like structures by arranging spines in R3 using directional forces. Each spine is attracted to its neighbor based on the temporal coincidence of afferent action potentials. The resulting spine location spatially encodes activation patterns. Proximity to the soma and to each other determines connectivity within an acyclic graph. Dendritic branches carry back-propagating action potentials used for credit assignment by Hebbian-style learning. Preliminary results suggest this method could be used to create spiking neural networks consisting of dendritic neurons capable of pattern recognition.