Unsupervised query-based learning algorithm and its application to Kohonen´s self-organizing maps

Query-based learning (QBL) algorithms have been introduced to supervised multilayer perceptrons to provide good training results for correct classification. Since their queried data are produced by external supervisors, conventional QBL methods can not be directly applied to unsupervised learning models which have no external supervisor exist. In this paper, a novel unsupervised QBL (UQBL) algorithm is proposed. In which, network´s training samples are further improved by both the goal-oriented selective attention and the self-regulation property. The proposed UQBL method considers not only the external stimulus but also the internal desire. It is not an anthropomorphic style that overestimates the importance of internal desire. The authors just try to combine two different system parameters for network training. This method is different from the conventional supervised/unsupervised learning algorithms. The authors´ experiments show that the proposed UQBL method can be successfully applied for Kohonen´s self-organizing maps (SOM). It can provide faster convergence and is more insensitive to network initialization than the standard SOM