Simplex ARTMAP: Building General Geometry Borders Among Predictions with Simplex-Shaped Classes

D. Gomes, M. Fernández-Delgado, and S. Barro (Spain)


ART neural networks. Geometric class shapes. Fuzzy ARTMAP. Simplex ARTMAP. Classification.


ART-based neural networks are popular algorithms for on line pattern recognition. Some of their drawbacks are re lated with their ability to learn complex borders among patterns belonging to different predictions. This may be due partially to the predefined geometrical shape of classes in ART networks (hyperboxes, hyperellipsoids, ...), which may limit their learning capabilities. This paper presents Simplex ARTMAP (SAM), an ART-based algorithm which uses simplex (hypertetrahedron)-shaped classes defined by selected training patterns, as opposed to classical ART net works, based on prototypes or similar. Input patterns in SAM are learnt by growing classes only in the direction of the input pattern, as opposite to existing ART-based net works, and without the need for the vigilance parameter. Results on two 2D datasets show, numerically and graphi cally, the ability of SAM to fit the class borders and suggest an interesting way to get higher performances using on-line learning.

Important Links:

Go Back