Discover and Cure: Concept-aware Mitigation of Spurious Correlation

• DOI: 10.48550/arxiv.2305.00650
• Publication Date: 2023-01-01
• Authors: Shirley Wu; Mert Yüksekgönül; Linjun Zhang; James Zou
• Keywords: Spurious relationship; Interpretability

Deep neural networks often rely on spurious correlations to make predictions, which hinders generalization beyond training environments. For instance, models that associate cats with bed backgrounds can fail predict the existence of in other environments without beds. Mitigating is crucial building trustworthy models. However, existing works lack transparency offer insights into mitigation process. In this work, we propose an interpretable framework, Discover and Cure (DISC), tackle issue. With human-interpretable concepts, DISC iteratively 1) discovers unstable concepts across different as attributes, then 2) intervenes data using discovered reduce correlation. Across systematic experiments, provides superior ability interpretability than approaches. Specifically, it outperforms state-of-the-art methods object recognition task a skin-lesion classification by 7.5% 9.6%, respectively. Additionally, theoretical analysis guarantees understand benefits trained DISC. Code are available at https://github.com/Wuyxin/DISC.