| 摘要 |
PROBLEM TO BE SOLVED: To provide a learning method capable of surely improving the generalizing capability of a neural network. SOLUTION: In a hierarchical neural network provided with an input unit, intermediate unit and output unit as an object, after the neutral network performs learning with the method of learning vector quantization (S1), pseudo- data to be equally distributed into respective class areas are generated by applying the method of balloon net model into the respective class areas defined in the result input space of that learning (S2) and while using these pseudo-data as a teacher vector, the weight of coupling between respective units is made to learn by the method of back propagation (S3). Since the pseudo-data are the values equally distributed in respective category areas, by performing learning through the method of back propagation with these data as teacher vectors, a learning convergent state is provided in an internal neutral network structure, having a statistically rational category identification boundary without depending on the initial value.
|
