PURPOSE: To provide an example learning method capable of surely performing example learning even with relatively less sample data.
CONSTITUTION: The minimum/maximum values of a feature amount axis where a supplied discrimination category is present are detected and a data range where the discrimination category is present is obtained (ST10), a prescribed margin width is set outside the data range, the prescribed number of pseudo different kind categories are prepared from feature amount data present in the area on the outside (ST11) and example learning (initial learning) is performed based on the pseudo different kind categories and the discrimination category (ST12 and 13). The learning is ended in a short time since the categories clearly different with each other are used and a rough boundary line is obtained. Then, the example learning is performed by using the discrimination category obtained by actually performing measurement and the different kind category (ST14 and 15). Since the result of the initial learning is used for the initial value of the start of the learning, the learning becomes the correction processing of the boundary line and convergence to an optimum solution is performed in a short time.
YAKURA TOKUMASA