In this paper, a Normalized Gaussian Network (NGnet) is introduced for online sequential learning that uses unit manipulation mechanisms to build the network model self-constructively. Several unit manipulation mechanisms have been proposed for online learning of an NGnet. However, unit redundancy still exists in the network model. We propose a merge mechanism for such redundant units, and change its overlap calculation in order to improve the identification accuracy of redundant units. The effectiveness of the proposed approach is demonstrated in a function approximation task with balanced and imbalanced data distributions. It succeeded in reducing the model complexity around 11% on average while keeping or even improving learning performance.
CITATION STYLE
Backhus, J., Takigawa, I., Imai, H., Kudo, M., & Sugimoto, M. (2016). Reducing redundancy with unit merging for self-constructive normalized Gaussian networks. In Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics) (Vol. 9886 LNCS, pp. 444–452). Springer Verlag. https://doi.org/10.1007/978-3-319-44778-0_52
Mendeley helps you to discover research relevant for your work.