Modular Network for Generalized Self-Organizing Map

• Published in: Denshi Jouhou Tsuushin Gakkai Kiso, Kyoukai Sosaieti fundamentals review Vol. 14; no. 2; pp. 97 - 106
• Main Author: TOKUNAGA, Kazuhiro
• Format: Journal Article
• Language: Japanese; English
• Published: Tokyo The Institute of Electronics, Information and Communication Engineers 01.10.2020; Japan Science and Technology Agency
• Subjects: Algorithms; Artificial neural networks; Dynamic models; Machine learning; modular network self-organizing map; Principal components analysis; Scientific visualization; Self organizing maps; self-organizing map; Topology; Visualization
• ISSN: 1882-0875; 1882-0875
• DOI: 10.1587/essfr.14.2_97

Data visualization is an important technique for interpreting complex data and finding relationships between data. Although principal component analysis is widely known as a method of data visualization, the Self-Organizing Map (SOM) proposed by Kohonen, one of the artificial neural networks, is also widely used as a data visualization tool. The SOM performs a topology-preserving transformation from a higher-dimensional vector space to a lower one, and generates a map that represents the relationships between data vectors. In some cases, however, it is necessary to generate maps based on the similarity between models that generate the data vectors. The author proposed a modular network self-organizing map (mnSOM: Modular Network SOM) as a method to solve such problems. The mnSOM has an architecture as a generalized SOM since the mnSOM can generate maps for a variety of models such as input-output functions, dynamical models, manifolds and so on. In this paper, the theory and learning algorithms of mnSOM are explained with an explanation of SOM.