Abstract
This paper presents a new method to improve generalization abilities for Support Vector Machines (SVM) based on topological data mapping used in Counter Propagation Networks (CPN). The proposal method produces new training data to be expanded or compressed while retaining topological data structure using competitive and neighbor algorithms on CPN. The number of new training data is controlled by the changing of units on a mapping layer of CPN. Using topological data mapping, interrupting data is created to sparse data regions and redundant data is removed from overlapping data regions. Especially, the characteristic of removing redundant data is connected to reduce the number of Support Vectors (SV) that treated for soft-margins. We applied our method to two classification datasets and a face dataset under illumination changes. The classification results indicate the basic characteristics of our method to be changed decision boundaries. The face recognition results show that our method provides not only improved generalization abilities, but also to be able to display spatial distributions of SV on a category map.
