PURPOSE: To obtain the leaning system of a neural network which can compose the neural network of high generalization ability by individually monitoring/ calculating the advance of the learning of an input pattern.

CONSTITUTION: A pattern P1 is inputted to the input layer 1 of the neural network (ST 110), and calculated in a forward direction so as to obtain an output pattern 05 at an output layer (ST 120). Next, the square error E1 of a teacher pattern T1 and the output pattern O1 is calculated (ST 130). Then, when the square error E1 is over a value α which is previously set, weight is updated in a reverse direction in accordance with back-propagation algorithm and when being less than the value α, weight is not updated (ST 140). The operation of ST 110 to 140 is executed to the other input patterns (P2 to PN) (ST 150) and when an average square error Em obtained by calculating to all the input patterns is less than set value β, the leaning of the neural network is stopped (ST 160, 170).