Concept Flow Models Improve Interpretability with Hierarchical Bottlenecks

This paper introduces Concept Flow Models (CFMs), an advancement over traditional Concept Bottleneck Models (CBMs) designed to improve interpretability and reduce information leakage in AI systems. CBMs project learned features into human-understandable concept spaces, but can suffer from information leakage when the number of concepts is high, leading to models exploiting spurious correlations. CFMs address this limitation by replacing the flat bottleneck with a hierarchical, concept-driven decision tree. Each node in this hierarchy focuses on a smaller, localized subset of discriminative concepts, allowing the model to progressively refine its prediction scope. The framework constructs these decision hierarchies from visual embeddings, distributes semantic concepts at each level, and trains differentiable concept weights via probabilistic tree traversal. Experimental results on various benchmarks demonstrate that CFMs maintain predictive performance comparable to flat CBMs while significantly reducing information leakage by limiting the effective concept usage. This hierarchical structure also provides transparent, auditable stepwise decision flows, particularly beneficial for tasks with inherent hierarchical class structures.