In SVM, we classify y based on whether f(x) > 0 or f(x) < 0.
I understand that in SVM with f(x) being linear in x, the support set is typically small (i.e., the number of support vectors is much smaller than the number of training examples).
My question is whether the same is still true when we do the kernel trick? In other words, when doing prediction with, say, RBF kernel, are we only using a small amount of examples?
I coded a SVM with RBF kernel myself and find this not to be the case (i.e., almost all examples are used for prediction). I wonder whether this should be the case.