Large-Margin Softmax Loss for Convolutional Neural Networks