Placing spline knots in neural networks using splines as activation functions.

Placing spline knots in neural networks using splines as activation functions.

Abstract

When using feed-forward neural networks with spline activation functions, the quality of approximation depends on the knot placement of spline functions. We demonstrate a method of choosing equidistant knots in each subdivision of the space when an arbitrary initial division is given, in order to keep the approximation error under a predefined limit.

Grafik Top
Authors
  • Schindlerova, Katerina
  • Verleysen, Michel
Grafik Top
Shortfacts
Category
Journal Paper
Divisions
Data Mining
Subjects
Kuenstliche Intelligenz
Journal or Publication Title
Neurocomputing
ISSN
0925-2312
Publisher
Elsevier
Page Range
pp. 159-166
Number
3-4
Volume
17
Date
1997
Export
Grafik Top