PURPOSE: To provide an improved neural network training method by which the character recognizing accuracy of the neural network can be improved.

CONSTITUTION: In a neural network training method, noise is simulated by using a modified training sequence using characters having widths thicker or narrower than those of normal characters. Preferably, the training of a neural network is first repeated by the prescribed number of times in accordance with a normal training sequence containing characters to be recognized, characters to be verified, and other characters. Then the training sequence is modified by making the width of the characters thicker or thinner. The modified sequence is applied along the sequence of characters having a normal width until the neural network is converged and excellent generalization and noise resistances are obtained.