A measure of how well different visual concepts can be distinguished in a model's learned feature space.