Hoeffding Concept Bottleneck Models with Applications to Overhead Images
Researchers have introduced Hoeffding Concept Bottleneck Models (HCBM), a novel approach to enhance the explainability of deep learning models in computer vision. Unlike existing methods that use linear aggregation of concept scores, HCBM employs non-linear and sparse aggregations based on Hoeffding functional decomposition. This method demonstrates robustness to inter-concept leakage and outperforms standard linear CBMs, showing particular promise in object detection tasks with overhead imagery. AI
IMPACT Introduces a more robust and explainable method for deep learning models, potentially improving trust and adoption in critical applications.