Interactive Concept Bottleneck Models

Concept bottleneck models (CBMs) are interpretable neural networks that first predict labels for human-interpretable concepts relevant to the prediction task, and then final label based on concept predictions. We extend CBMs interactive settings where model can query a human collaborator some concepts. develop an interaction policy that, at time, chooses which request so as maximally improve prediction. demonstrate simple combining uncertainty influence of achieves strong performance outperforms static approaches well active feature acquisition methods proposed in literature. show CBM achieve accuracy gains 5-10% with only 5 interactions over competitive baselines Caltech-UCSD Birds, CheXpert OAI datasets.