The work “Deferring Concept Bottleneck Models: Learning to Defer Interventions to Inaccurate Experts” featuring Riccardo Massidda, Francesco Giannini, Pietro Barbiero, Mateo Espinosa Zarlenga, Roberto Pellungrini, Gabriele Dominici, Fosca Giannotti and Davide Bacciu has been accepted at NeurIPS 2025!
This paper proposes a novel approach to enhance Concept Bottleneck Models (CBMs) by enabling them to defer interventions to human experts, even when those experts may be inaccurate. For this purpose, we introduce Deferring CBMs (DCBMs). DCBMs learn to decide when to defer concepts and task predictions based on the reliability of the experts. Following standard Learning to Defer literature, we show how our approach can be derived in a principled manner and propose a surrogate loss that is Bayes-consistent under the independence assumption of concepts. We finally showcase the effectiveness of DCBMs in our experimental evaluation in addressing a few shortcomings of standard CBMs.
A final thanks to all the amazing Team.
See you in San Diego ⛱️🌊!