<ul><li>Concept bottleneck models (CBMs) aim to decompose predictions into human interpretable concepts.</li><li>Annotations used for training CBMs are often noisy, impacting prediction performance and interpretability.</li><li>A study on noise in CBMs shows that corruption impairs prediction performance, interpretability, and intervention effectiveness.</li><li>A proposed two-stage framework aims to mitigate vulnerability by stabilizing learning of noise-sensitive concepts during training and correcting uncertain concepts during inference.</li></ul>

An Analysis of Concept Bottleneck Models: Measuring, Understanding, and Mitigating the Impact of Noisy Annotations

Discover more