HCBM: Non-linear Concept Bottleneck Models Beat Linear CBM in Overhead Imaging
New explainable AI method reduces concept leakage using Hoeffding functional decomposition.
Explainability remains a critical bottleneck for deploying deep learning in high-stakes computer vision applications. Concept Bottleneck Models (CBM) have emerged as a promising approach, using a bottleneck of high-level concepts to provide interpretable predictions. However, conventional CBM methods rely on linear aggregation of concept scores, which often requires many concepts—undermining interpretability and allowing information leakage between concepts. The underlying relationship between concepts and output logits is rarely linear.
To address these limitations, researchers present Hoeffding Concept Bottleneck Models (HCBM). HCBM leverages the Hoeffding functional decomposition of gradient-boosted trees to learn non-linear yet sparse aggregations of concept scores. This results in compact predictions using prime implicants, which are inherently more interpretable. Extensive experiments demonstrate HCBM's robustness to inter-concept leakage and superior performance over standard linear CBM on both classification and object detection tasks. Notably, HCBM excels on challenging overhead satellite imagery, where complex spatial relationships make interpretability particularly valuable.
- HCBM replaces linear concept aggregation with non-linear, sparse Hoeffding decomposition of gradient-boosted trees.
- The method reduces inter-concept information leakage, a known weakness of standard linear CBM.
- Achieves state-of-the-art performance on overhead image classification and object detection benchmarks.
Why It Matters
Makes high-stakes vision AI more trustworthy and accurate, especially for satellite and aerial imagery analysis.