Figure 3.

Summary of the classification results. A, Four facial features that were measured for classification analysis. B, A linear SVM classifier that was trained to distinguish negative versus positive surprised faces demonstrated 75% accuracy from cross-validation testing. Error bars indicate SEM. C, Absolute values of the feature weights indicate that the mouth regions have more informational value than the eye regions for this trained classifier. D, An additional validation test was performed on a holdout set of surprised faces that were never used for the cross-validation training of the classifier, and showed a classification accuracy of 87.5%.