Chi-square Loss for Softmax: an Echo of Neural Network Structure