Multiple-Instance Learning From Distributions