The robustness of SVM is actually due to its slightly more sophisticated
characteristics than this large margin view might suggest. In particular,
notwithstanding the largemargin reliability of the classifier, the
algorithmcan still be sensitive to outliers, but SVMs offer the chance to
control the risk of overfitting, also indicated as bias-variance trade off
(Mountrakis et al., 2011), through the tuning of one parameter, called
C. If the value of C is set very large, then the learned hypothesis may
fit the training set very well, but it will fail to generalize to new examples
(predict classes on new pixels); this can be translated also as a
hypothesis with higher variance and lower bias.