Stabilizing competitive learning during on-line training with an anti-Hebbian weight modulation

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.