My 2c From personal experience if your independant variable is categorical or discrete in nature then linear models work well. In contrast when your dependant variables are continous or you are using svm for regression rbf seems to work. Svm are really good algorithms when you have limited data size. For large data it becomes computationally expensive

I fundamentally doubt the ability of neural nets to generalize without using large amount of power and compute. Biochips will be critical to make generalized learners.