An Online Self-constructive Locally Updated Normalized Gaussian Network with Localized Splitting

In sequential learning schemes, dynamic model adaptation is preferable over static model size selection. In this paper, we apply an incremental model selection approach to a locally updated Normalized Gaussian network (NGnet), and improve it for better robustness against negative interference. Model adaptation is enabled by applying some unit manipulation mechanisms, including a produce, delete and split manipulation, to increase or decrease model complexity. Here, split uses a static threshold for its manipulation, and we suggest a dynamic thresholding approach that selects a threshold according to local information. In our experiments, the NGnet is tested for a function approximation task with balanced and imbalanced sample distributions. We compared local thresholding to static and dynamic global thresholding, and results show that localized thresholding improves performance for both test cases, and especially for imbalanced sample data. Therefore, localized splitting is preferable especially in test cases where negative interference is likely.