Skip to Main Content
This chapter proposes a new mechanism named the self-organizing tree map (SOTM). The motivation of the new method is twofold: (a) to keep the ART's ability to create new output neurons dynamically while overcoming the Global threshold setting problem; (b) to keep the SOM's property of topology preservation while strengthening the flexibility of adapting to changes in the input distribution and maximally reflecting the distribution of the input patterns. In the SOTM, relationships between the output neurons can be dynamically defined during learning. There are thus, two different levels of adaptation in the SOTM, which involves weight adaptation and structure adaptation. The basic principle underlying competitive learning is vector quantization. The chapter demonstrates dynamic topology and classification capability are two prominent characteristics of the SOTM. Finally, the SOTM model not only enhanced the ART's autonomous category classification and the SOM's topology preservation, but also overcame their weaknesses.