NEURAL NETWORKS IN NON-EUCLIDEAN SPACES.

Wlodzislaw Duch, Rafał Adamczak, Geerd H.F. Diercksen,
Department of Informatics, Nicolaus Copernicus University,
Grudziadzka 5, 87-100 Torun, Poland & MPA.

Neural Processing Letters 10 (1999) 201-210




Multilayer Perceptrons use scalar products to compute weighted activation of neurons. Transforming the input vectors into an extended space in which non-Euclidean metric is used for normalization has dramatic effect on the shapes of decision borders. An illustrative example showing these changes is provided.

Paper in PDF, 782 KB

Projects on similar subject and BACK to the on-line publications of W. Duch.