Sparse approximation through boosting for learning large scale kernel machines