Using the Geometrical Distribution of Prototypes for Training Set Condensing