Walking the Web of Concept-Class Relationships in Incrementally Trained Interpretable Models