Modular network SOM : The architecture, the algorithm and applications to nonlinear dynamical systems

Proceedings of 5th Workshop on Self-Organizing Maps （WSOM05）, 2005年9月5日, Paris, FranceThis paper proposes a generalized framework ofSOM applicable to more extended data classes rather than vector data. The generalization is realized by adopting the idea of a modular network; thus it is called a modular network SOM (mnSOM), in which each nodal vector unit in a conventional SOM is replaced by a function module such like a neural network or a trainable algorithm. First, we introduce the basic idea ofan mnSOM, and then byfocusing on a class ofmnSOM that consists of multi-layer perceptron modules, the architecture and the algorithm are presented. Finally, some applications to nonlinear dynamical systems are reported