Separating different factors of variation (like expression and identity) in a model's learned representations.