if the learner doesn't output class posteriori, a simple kernel density estimator is fitted first. A bit awkward maybe.