PURPOSE: To promptly decide the number of intermediate layer units by using a rank value of output covariance matrix of intermediate layer as an intermediate layer unit number.

CONSTITUTION: The unit numbers of input layer and output layer are preset, and the unit numbers of an intermediate layer is assumed to be that of an initial unit so that a neural net is configured. Then, learning is performed by using back propagation (S2), and after convergence (S3), the rank of an intermediate layer output covariance matrix is decided (S4) by singular value resolving method(SVD). Using the obtained intermediate layer unit numbers, learning is performed by back propagation (S5) far convergence (S6). The unit number of intermediate layer is decided to the smaller rank in the rank of a weight matrix between the input layer and the intermediate layer obtained or the rank of an input pattern covariance matrix.