Core Concepts
Concept Bottleneck Models can be significantly improved by leveraging concept relations to realign concept assignments after human interventions.
Abstract
The paper proposes a concept intervention realignment module (CIRM) to address the independent treatment of concepts during test-time interventions in Concept Bottleneck Models (CBMs). This independent treatment limits the efficacy of human interventions, as correcting one concept does not influence the use of other related concepts.
The CIRM consists of two components: a concept realignment model (CRM) that updates concept predictions based on intervened concepts, and an intervention policy that selects the next concept to intervene on. The CRM is trained to leverage concept relations and update the concept assignments accordingly.
Experiments on three standard benchmarks (CUB, CelebA, AwA2) show that the proposed CIRM can significantly improve the efficacy of human interventions. Across both concept prediction accuracy and overall classification accuracy, performance increases more rapidly with interventions compared to a baseline without concept realignment. In some cases, the number of interventions needed to reach a target performance is reduced by over 70%.
The CIRM can be applied as a post-hoc module on top of existing CBM approaches, as well as integrated into the training process of intervention-aware models. The results demonstrate the versatility and practical relevance of the proposed concept realignment approach in facilitating human-model collaboration and enabling the deployment of CBMs in resource-constrained environments.
Stats
"Concept prediction loss can be reduced by a factor of 10 with half the number of interventions."
"Classification accuracy can reach the upper-bound with 50% fewer interventions using concept realignment."
Quotes
"Concept realignment can significantly improve intervention efficacy; significantly reducing the number of interventions needed to reach a target classification performance or concept prediction accuracy."
"Our experiments provide strong evidence that concept intervention realignment is crucial to best leverage human feedback in concept-based decision systems."