This Paper describes the implementation methods of MLP (mulilayer perceptrons) neural networks to recognize or reject handwritten numerals and non-nummerals. The MLP has known to be a very efficient classifier to recognize handwritten numerals in terms of recognition accuracy, speed, and memory requirements. In the previous researches, however, researchers have focused on the only numeral inputs and have not payed attention to the non-numeral inputs with respect to recognition accuracy, rejection rates, and other characteristics. In this paper, we present some implementation methods of the MLP in the environments that numeral and non-numerals are mixed. The MLPs have been developed by three methods, and investigated with three error types introduced. The experiments have been conducted on a total of 66,701 images of numerals and non-numerals. The promising method to recognize numerals and reject non-numerals has been described in terms of the three error types.