Training models to learn p(x,y) the full joint probability of features and targets, rather than just conditional p(y|x).