ML Interview Q Series: Why might you use an energy-based model’s cost function (e.g., contrastive divergence) in certain generative tasks, and how does it compare to typical likelihood-based losses?