Learning method, information conversion device, and recording medium

An information processing device for a biometric authentication comprising:

• a data pair generator configured to generate, using different generation methods, data pairs from feature amount vectors included in a learning data set;
• a counter configured to count any one of or some of a number of labels added to each of the feature amount vectors included in the learning data set, a number of types of the labels, a number of feature amount vectors added with a same label, and a number of data pairs used for learning of a hyperplane;
• a data pair generation method selector configured to select, according to a result of the counting, one or more generators that generate the data pairs from the learning data set based on a ratio of the feature amount vectors added with one label, a ratio according to the number of the types of the label, a ratio according to the number of the feature amount vectors added with the same label, and a ratio according to the number of the data pairs used in learning of the hyperplane;
• a calculator configured to calculate, , using a result of the counting, a ratio of a number of the data pairs to be generated by each of the selected one or more generators to a total number of the data pairs to be used,
• a data pair selector configured to select, at the calculated ratios, the data pairs generated by the selected one or more generators;
• and
• a learner configured to learn, using the selected data pairs selected by the data pair selector, the hyperplane that divides a feature amount vector space.

A non-transitory computer-readable recording medium storing a learning program that causes a computer to execute a biometric authentication comprising:

• counting any one of or some of a number of labels added to each of feature amount vectors included in a learning data set, a number of types of the labels, a number of feature amount vectors added with a same label, and a number of data pairs used for learning of a hyperplane;
• first selecting, according to a result of the counting, one or more data pair generation methods from a plurality of previously stored data pair generation methods that generate the data pairs from the learning data set based on a ratio of the feature amount vectors added with one label, a ratio according to the number of the types of the label, a ratio according to the number of the feature amount vectors added with the same label, and a ratio according to the number of the data pairs used in learning of the hyperplane;
• calculating, using a result of the counting, a ratio of a number of the data pairs to be generated by each of the selected one or more generation methods to a total number of the data pairs to be used;
• generating, using the selected one or more generation methods, the data pairs from the feature amount vectors included in the learning data set;
• second selecting, at the calculated ratios, the data pairs generated by the selected one or more generation methods;
• and
• learning, using the generated data pairs, the hyperplane that divides a feature amount vector space.