Abstract of FKI-183-93
Title: Theoretical Issues Concerning the Representation of
Continuous-Valued Input and Output Data in Neural
Autors: Margit Kinder
Revision Date: June, 1993
Category: Technical Report (Forschungsberichte Künstliche Intelligenz)
Abstract: There exist all kinds of problems where both input and output
data for neural networks are continuous and vector-valued.
From our previous works we know that distributed representa-
tions of the input data are extremely useful for neural net-
works to embody good generalization skills and also to model
forbidden regions in data space. Units are only located in
those regions where data has occurred in the learning process.
In this paper, we analyze Kohonen's self-organizing feature
maps with respect to distributed representations and in this
respect come up with a comparison to Radial Basis Functions.
Concerning the output data, we give two interpretations for
distributed representations. First, the center of
gravity interpretation for which we explain some severe
problems. Second, the pseudo-inverse of the matrix defined by
the positions of the representation units.
Keywords: self-organizing feature maps; function approximation;
pseudo-inverse; normalization of input vectors;
topologically distributed representation; Radial Basis
Size: 8 pages
Copyright: The ``Forschungsberichte Künstliche Intelligenz''
series includes primarily preliminary publications,
specialized partial results, and supplementary
material. In the interest of a subsequent final
publication these reports should not be copied. All
rights and the responsability for the contents of the
report are with the authors, which would appreciate