Hoeffding Concept Bottleneck Models with Applications to Overhead Images

Researchers introduce Hoeffding Concept Bottleneck Models (HCBM), a novel approach to explainable AI that uses non-linear aggregation of concept scores instead of traditional linear methods. The technique demonstrates improved performance on classification and object detection tasks while maintaining robustness against information leakage between concepts.

This research addresses a fundamental challenge in explainable AI: creating models that are both interpretable and accurate without sacrificing performance. Traditional concept bottleneck models rely on linear combinations of learned concepts to make predictions, but this approach requires numerous concepts to capture complex relationships, paradoxically reducing explainability. HCBM solves this by leveraging Hoeffding functional decomposition from gradient-boosted trees to discover non-linear, sparse aggregations of concepts that compress information into prime implicants—a more compact and human-interpretable representation.

The advancement matters significantly for high-stakes applications like medical imaging, autonomous vehicles, and satellite image analysis where decisions must be justified to regulators and stakeholders. By reducing the number of active concepts while maintaining or improving accuracy, HCBM makes model reasoning more transparent and harder to exploit through concept-level attacks that attempt to inject spurious correlations.

The focus on overhead imagery demonstrates practical applicability to real-world computer vision challenges. The technique's robustness to inter-concept leakage—where concepts unintentionally encode information about each other—addresses a known vulnerability in existing explainable AI systems. This matters for industries deploying AI in regulated environments where concept drift and model manipulation pose risks.

Future developments will likely explore how HCBM scales to very large concept sets and whether the approach generalizes to other domains beyond vision. The research also suggests integration opportunities with reinforcement learning systems that require interpretable decision pathways for safe deployment.

• →HCBM uses non-linear concept aggregation to improve both explainability and accuracy over linear baseline models
• →The approach reduces required concept count while maintaining or exceeding classification performance
• →Hoeffding decomposition provides robustness against inter-concept information leakage attacks
• →Extended applicability to object detection with overhead imagery demonstrates practical deployment potential
• →Compact prime implicant representations enhance human interpretability of model decisions