Chaos-based Learning
Paul F. M. J. Verschure
Institute for Informatics, University of Zurich,
Winterthurerstrasse 190, CH-8057 Zurich, Switzerland
Abstract
It is demonstrated that the chaotic properties of neural networks (in this case networks defined by the generalized delta procedure) can be used to improve their learning performance. By adaptively varying the learning-rate parameter an annealing mechanism can be introduced that is founded in chaos. The proposed mechanism, chaos-based learning, provides faster convergence than standard back-propagation and also seems to provide a computationally less intensive alternative to other back-propagation accelerating techniques by using adaptive step-size control.
