Title: Hoeffding Concept Bottleneck Models with Applications to Overhead Images

Abstract:

The capacity to interpret deep learning algorithms is essential for computer vision systems involved in critical decision-making processes. Concept Bottleneck Models (CBMs) have emerged as a promising solution for classification tasks, delivering both high accuracy and interpretability by leveraging a bottleneck of high-level concepts. However, conventional CBM approaches typically depend on linearly aggregating concept scores to derive predictions. This reliance on linearity often necessitates the use of numerous concepts, which can compromise interpretability and increase the risk of information leakage. Furthermore, the relationship between underlying concepts and output logits is rarely linear in nature.

To address these limitations, we propose Hoeffding Concept Bottleneck Models (HCBM). This framework utilizes the Hoeffding functional decomposition of gradient-boosted trees to achieve non-linear and sparse aggregation of concept scores. By employing prime implicants, HCBM generates concise and compact predictions. Our theoretical analysis demonstrates that HCBM is robust against interconcept leakage, and empirical results from extensive experiments confirm that it surpasses standard linear CBMs in performance. While initially designed for classification, the HCBM architecture is adaptable to object detection tasks. We illustrate its superior capabilities in this domain by applying it to the challenging scenario of overhead image analysis.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC