Hierarchical, Interpretable, Label-Free Concept Bottleneck Model
New model outperforms state-of-the-art CBMs in accuracy while providing multi-level explanations without manual labels.
A team of researchers, including Haodong Xie, Yujun Cai, and Federico Tavella, has introduced HIL-CBM, a novel Hierarchical Interpretable Label-Free Concept Bottleneck Model. Published on arXiv, this work tackles a core limitation of existing Concept Bottleneck Models (CBMs), which are designed to make AI decisions interpretable by predicting labels through human-understandable concepts. Current CBMs operate at a single semantic level, unlike human cognition, which identifies objects using a hierarchy of features from general to specific. HIL-CBM bridges this gap by structuring both concepts and predictions across multiple abstraction levels, all without requiring manually annotated concept relationships—a significant data efficiency win.
The model's architecture is built on two key technical innovations. First, it employs a gradient-based visual consistency loss that encourages different abstraction layers to focus on similar spatial regions in an image, ensuring coherent multi-level reasoning. Second, it trains dual classification heads that operate on feature concepts at different abstraction levels, allowing predictions and their corresponding explanations to be aligned from abstract (e.g., 'vehicle') to concrete (e.g., 'sedan'). In benchmark tests, HIL-CBM demonstrated superior classification accuracy compared to state-of-the-art sparse CBMs. Crucially, human evaluators found the hierarchical explanations provided by HIL-CBM to be more accurate and interpretable, validating its approach to mimicking human-like reasoning.
This research represents a meaningful step forward in explainable AI (XAI), particularly for high-stakes fields like medical imaging or autonomous systems where understanding the 'why' behind a model's decision is as critical as the decision itself. By making AI explanations more natural and aligned with human cognitive processes, HIL-CBM could lower the barrier for experts to trust and effectively collaborate with complex vision models.
- Eliminates need for manual concept annotations, using a label-free, hierarchical approach to feature learning.
- Outperforms state-of-the-art sparse Concept Bottleneck Models (CBMs) in classification accuracy on benchmark datasets.
- Human evaluations confirm the model provides more interpretable and accurate multi-level explanations than previous methods.
Why It Matters
Enables more trustworthy AI for critical applications by providing human-aligned, hierarchical explanations without costly data labeling.