Multiple Kernel Learning: A Unifying Probabilistic Viewpoint