Hypercolumn model: A modified model of neocognitron using hierarchical self-organizing maps

Abstract

The sequential recognition method using the selective attention mechanism of Neocognitron (NC) is realizable and very effective for images of multiple targets. NC, however, is not applicable to general (large-scale, gray or color complex object) images, because of both a low reduction rate of coding scheme and an initial state dependency of competitive learning. In this paper, "hypercolumn model (HCM)" is introduced to modify NC and to apply the sequential recognition to general images. Two properties of the hierarchical self-organizing maps proposed by Lampinen: a coding scheme using topographic mapping and an initial state independence of learning resolve the two disadvantages of NC, respectively. Experimental results will show that a recognition accuracy of HCM for general images was from 10 percent to 30 percent higher than one of NC, and a sequential recognition performed well.