Dynamic Deletion of Units for Self-Organizing Map by Introducing a New Measure of Unit's Contribution to Learning
Eiji Uchino, Masaki Kawamura, Nagata Kazuhiro · Journal of Japan Society for Fuzzy Theory and Intelligent Informatics · 2002
This paper proposes a dynamic deletion method of units for SOM (Self-Organizing Map) by using the updated information of each unit on the competing layer. Concretely, during the process of learning, a unit with a low contribution to learning, is removed dynamically by each unit's ability to judge for itself (extinction of unit). This process may imitate a cell death in a living body. This dynamic deletion of units enables SOM to map the input data on the competing layer accurately with a small number of units. Furthermore, by using the proposed method, the input space with discontinuous distribution of data can be learned correctly.