Toward a Characterization of Loss Functions for Distribution Learning