Improving the perforce of differential competitive learning model in clustering tasks
Guilherme A. Barreto, A.F.R. Araujo · 1998
In this paper we propose two neural algorithms that can be considered a simplification and a generalization of the Differential Competitive Learning (DCL) neural network, respectively. Firstly, we suggest some simplifications for the original DCL model to eliminate redundant aspects of the competition mechanism. We get rid of the lateral connections arguing that it is possible because the winning neuron is chosen based solely on metrical similarity measures and the lateral feedback weights play no effective role. The activation rule is made simpler requiring less computational effort. In the second model, we show how to combine lateral connections with metrical relations on the activation and the learning rules of DCL to effectively estimate cluster centroids. This model is also less sensitive to weight initialization. A number of simulations are carried out to compare the presented models in unsupervised clustering tasks.