PURPOSE: To provide a learning system for a neural network capable of improving classification accuracy for unknown data even when a noise is included in learning data.

CONSTITUTION: This system is provided with a classification rule execution means 1 which inputs the feature quantity vector 12 of the learning data, and outputs a known rule classification result vector and a classification result reliability, a distribution mean calculation means 2 which calculates the distribution density of the feature quantity vector 12 and the mean classification result vector of the learning data, a teacher vector calculation means 3 which outputs a teacher vector 15 in which weighing is applied to the classification result of the learning data or that by a known rule, and a teacher data storage means 4 which holds plural pieces of teacher data consisting of a pair of feature quantity vector 14 and teacher vector 15 and supplies them to a coupling load adjusting means 16.