Improving model performance via concept realignment

Paper Title:

Improving Intervention Efficacy via Concept Realignment in Concept Bottleneck Models

Published on:

2 May 2024

Primary Category:

Machine Learning

Paper Authors:

Nishad Singhi,

Jae Myung Kim,

Karsten Roth,

Zeynep Akata

Bullets

Key Details

•

Proposes concept intervention realignment module to leverage concept relationships

•

Finds independent concept treatment reduces intervention efficacy

•

Realignment module updates related concepts after intervention

•

Significantly reduces interventions needed for target performance

•

Easily integrates into existing concept-based models

Explore the topics in this paper

concept-interventions

concept-relationships

human-interventions

image-understanding

intervention-modules

AI generated summary

Improving model performance via concept realignment

This paper proposes a concept intervention realignment module to improve the efficacy of human interventions in concept bottleneck models. It finds that existing approaches often require numerous costly human interventions per image. This is driven by independent treatment of concepts during intervention. To address this, the paper introduces a module to realign concepts based on their relationships, significantly reducing required interventions.