A Comparison between Growing and Variably Dense Self Organizing Maps for Incremental Learning in Hubel Weisel Models of Concept Representation

Abstract

Hubel Weisel models of pattern recognition are thought to be anatomically and physiologically faithful models of information representation in the cortex. They describe sensory information as being encoded in a hierarchy of increasingly sophisticated representations across the layers of the cortex. They have been shown in previous studies as robust models of object recognition. In a previous work, we have also shown Hubel Weisel models as being capable of representing a hierarchy of concepts. In this paper, we explore incremental learning with respect to Hubel Weisel models of concept representation. The challenges of incremental learning in the Hubel Weisel model are discussed. We then compare the use of variably dense self organizing maps to perform incremental learning against the original implementation using growing self organizing maps. The use of variable density self organizing maps shows better results in terms of the percentage of documents correctly clustered. The percentage improvement in clustering accuracy is in some cases up to 50% over the original GSOM implementation for the incrementally learnt module. However, we also highlight the issues that make the evaluation of such a model a challenging one.