Abstract: Training sets usually contain large amount of similar samples, resulting in a longer training time and poor performance. To deal with this problem, a training sample selection method for neural networks based on nearest neighbor (NN) rule was proposed. Considering the significance of train sets for the performance of neural networks, the proposed method combined simplicity of nearest neighbor (NN) with high accuracy of neural networks and utilized the modified NN rule to select the most representative samples as a new training set. Experimental results show that the presented method can eliminate the redundancy, achieve higher recognition accuracy and better generalization ability with fewer samples and less training time.