Concept Bottleneck Model with Zero Performance Loss

Abstract

Interpreting machine learning models with high-level, human-understandable concepts has gained increasing importance. The concept bottleneck model (CBM) is a popular approach for providing such explanations but typically sacrifices some prediction power compared with standard black-box models. In this work, we propose an approach to turn an off-the-shelf black-box model into a CBM without changing its predictions or compromising prediction power. Through an invertible mapping from the model’s latent space to a concept space, predictions are decomposed into a linear combination of concepts. This provides concept-based explanations for the complex model and allows us to intervene in its predictions manually. Experiments across benchmarks demonstrate that CBM-zero provides comparable explainability and better accuracy than other CBM methods.