Measuring and Addressing Information Leakage in Concept Bottleneck Models

Concept Bottleneck Models (CBMs) are transparent models that use a human-interpretable bottleneck layer to improve interpretability in tasks such as image classification. However, model transparency does not guarantee that decisions are supported by relevant information. When this is not the case, we speak of information leakage. To provide a quantitative measure of this property, we introduce a metric called Irrelevant Concept Contribution (ICC), the first metric that directly quantifies information leakage for concept bottleneck models. Using this metric, we propose a comprehensive comparison of bottleneck architectures, training strategies, and decision processes. In particular, we demonstrate that state-of-the-art performance in image classification can be achieved using only relevant concepts that contribute positively to the decision process on commonly used datasets. This makes the decision process relevant by design and in an explicit way.