Can dimensionality reduction improve the performance of an SVM, and if so, under what circumstances?