This paper presents a new approach to identify the activity of input attributes efficiently in the wrapper model of feature selection. The relevant features are selected by the diversity among the inputs of the neural network and the entire process is done depending on several criteria. While the most of existing feature selection methods use all input attributes by examining network performance, we use here only the attributes having relatively high possibilities to contribute to the network performance knowing preceding assumptions. The proposed diversity-based feature selection method (DFSM) can therefore significantly reduce the size of hidden layer priori to feature selection process without degrading the network performance. We tested DFSM to several real world benchmark problems and the experimental results confirmed that it could select a small number of relevant features with good classification accuracies.