PURPOSE: To prepare a precise reference pattern by adding a new reference pattern when the distance (d) from an input X is larger than the threshold value E even with a reference pattern having the least distance (d).

CONSTITUTION: A pattern matching part 4 calculates the distances between all reference patterns Y and an input X and outputs the category name of a reference pattern having the least distance (d) between the pattern Y and the input X as the recognizing result. The distance (d) is calculated as d= X - Y2/z. Then a reference pattern updating part 5 calculates the distances between all reference patterns and the input X to an input vector X whose category (m) is decided. When the least distance value dm is larger than the threshold value E set previously, an input pattern X is registered as a new reference pattern together with registration of a constant (Z). As a result, a precise reference pattern is obtained even though the learning vectors are sequentially fetched.