Stabilizing Competitive Learning during On-line Training with an
Anti-Hebbian Weight Modulation
S. Tavitian, T. Fomin and A. Lõrincz
Proceedings of ICANN'96, Bochum
pp. 791--796
(1996)
Abstract
Competitive learning algorithms are statistically driven schemes
requiring that the training samples are both representative and randomly
ordered. Within the frame of self-organization, the latter condition
appears as a paradoxical unrealistic assumption about the temporal
structure of the environment. In this paper, the resulting vulnerability
to continuously changing inputs is illustrated in the case of a simple
space discretization task. A biologically motivated local anti-Hebbian
modulation of the Hebbian weights is introduced, and successfully used
to stabilize this network under real-time-like conditions.
