An arithmetic unit 15 multiplies a codeword X and a correlation matrix W. A comparison unit 17 compares the products output from the arithmetic unit 15 with a threshold .+-.TH. A learning state monitor 18 updates the correlation matrix based on the comparison results from the comparison unit. When all of the codewords are updated and the calculation results using the updated correlation matrix fall within a predetermined range, the threshold is changed to repeat the learning operation again for all of the codewords. When the calculation results obtained from the updated correlation matrix do not fall within a predetermined range and equal the previous calculation results, the update value for the correlation matrix is decreased to repeat the learning operation again for all of the codewords. When the update value becomes 0, an optimal threshold as well as an optimal correlation matrix are obtained, thereby completing the learning operation.