Only a small number of multipliers have nonzero values and they are associated with the
so-called support vectors, which form the boundaries of the classes. The maximal margin
classifier can be generalized to nonlinearly separable data via two approaches. One is to introduce
a soft margin parameter C to relax the constraint that all the training vectors of a certain class lie
on the same side of the optimal hyperplane. This approach is effective in case of noisy data. The
other is to transform input vectors into a higher dimensional feature space by a map function