Mitigating Reliance on Spurious Correlations in Deep Neural Classifiers without Annotation

A self-guided framework that automatically detects and mitigates a classifier's reliance on spurious correlations in the data without requiring any prior annotations.