One reason that SVM often performs better than ANN is that SVM is designed to minimize structural risk whereas ANN are usually based on minimization of empirical risk, that is the minimization of the number of misclassified points on the training set.