Rapid Training for Self-Organizing Neural Networks with Incremental

Abstract

Kohonen's self-organizing maps (KSOM) can generate mappings from high-dimensional pattern spaces to lower dimensional topological structures. The main features of this kind of mappings are the formation of topology preserving maps. To overcome some limitations of KSOM, self-organizing neural networks with incremental learning (SONNIL) can be used. SONNIL can change their topological structures during learning. Two kinds of SONNIL model were present by B. Fritzke, i.e., growing cell structures (GCS) and growing neural gas (GNG). To speed up the training for SONNIL, based on GCS and GNG, we present two SONNIL variants, multiple GCS and double GNG. This paper first gives an introduction to KSOM and neural gas networks. Then, we discuss GCS and GNG models. Our multiple GCS and double GNG are present in the section 4. It is ended with some testing comparison and conclusions