Gaussian Radial and Kernel Networks with Varying and Fixed Widths

Authors: Věra Kůrková
Publication year: 2013
Publication type: Konferenční příspěvek (mezinárodní konf.)
Periodical: SOFSEM 2013: Theory and Practice of Computer Science
Projects: Learning of functional relationships from high-dimensional data
Bibliographic citation: KŮRKOVÁ, Věra. Gaussian Radial and Kernel Networks with Varying and Fixed Widths. In: VAN EMDE BOAS, P., GROEN, F.C.A., ITALIANO, G.F., NAWROCKI, J., SACK, H., eds. SOFSEM 2013: Theory and Practice of Computer Science. Vol. 2. Prague: Institute of Computer Science AS CR, 2013, S. 95-102. ISBN 978-80-87136-15-7.
ID: 389193

Abstract

The role of widths of Gaussians in computational models which they generate is investigated. Suitability of Gaussian kernel models with fixed widths for regression is proven in terms of their universal approximation capability. Large sets of argminima of error functionals minimized during learning from data over Gaussian networks with varying widths are described. Dependence of stabilizers modelling generalization on widths of Gaussian kernels and the input dimension is estimated.