Contrast multiple facial expression recognition experiments and found that using SVM instead of softmax layer can achieve better classification results(65.47% accuracy on fer2013 dataset).