A novel technique for unsupervised learning in feature space is presented. The feature space is considered as composed of two distinct sources

“mode”and“valley”

in the point of view of information theory. An entropy-based threshold is taken to distinguish the discrete cells in the feature space. The cells labeled as“mode”are then chained to form mode areas. Thereafter a modified Akaike’s information criterion is proposed to solve the cluster validity problem. After all the parameters are estimated

a labeling algorithm is developed based on the majority game theory. The method is applied to color image segmentation. The segmentation process is completely autonomous.