Concept Flow Models: Anchoring Concept-Based Reasoning with Hierarchical Bottlenecks

The authors introduce Concept Flow Models (CFMs), a novel architecture that enhances interpretability in machine learning by utilizing a hierarchical decision tree structure instead of a flat bottleneck in Concept Bottleneck Models (CBMs). CFMs organize discriminative concepts into a decision hierarchy, allowing for localized focus and mitigating information leakage as the number of concepts increases. Extensive benchmark experiments show that CFMs maintain predictive performance comparable to flat CBMs while providing transparent reasoning through stepwise decision flows, which is crucial for practitioners seeking to build interpretable AI systems.