Learning Exceptional Subgroups by End-to-End Maximizing KL-divergence