抄録
In this paper, we exhibit that partition errors are equivalent to each other when the numbers of inputs in a partition space are mutually equal, and describe that the average distortion is asymptotically minimized. According to the equinumber principle, a creation method of competitive learning is proposed with the objective of avoiding the initial dependency of reference vectors. The present approach which has output units without neighboring relations equalizes the numbers of inputs in a partition space. To begin with, only one output unit is prepared at the initial stage, and a reference vector, according to the unit, is updated under competitive learning. Then output units are created sequentially to reach a prespecified number of neurons based on the equinumber principle, and competitive learning occurs until the termination condition is satisfied. Experimental results show the effectiveness of the present technique in the average distortion. Furthermore the present approach is applied to image data and the validity in employing as an image coding system is examined.