Towards Better Generalization and Interpretability in Unsupervised Concept-Based Models