Identifying Spurious Correlations and Correcting them with an Explanation-based Learning

TL;DR

This paper introduces a method to detect and correct spurious correlations in image classification models by analyzing prediction certainty changes with image perturbations, enhancing model trustworthiness.

Contribution

It proposes a simple, effective approach to identify spurious correlations and an explanation-based learning method to remove them from trained models.

Findings

Successfully identified overdependence on spurious regions

Demonstrated removal of spurious correlations improves model reliability

Applicable to datasets with synthetic spurious regions

Abstract

Identifying spurious correlations learned by a trained model is at the core of refining a trained model and building a trustworthy model. We present a simple method to identify spurious correlations that have been learned by a model trained for image classification problems. We apply image-level perturbations and monitor changes in certainties of predictions made using the trained model. We demonstrate this approach using an image classification dataset that contains images with synthetically generated spurious regions and show that the trained model was overdependent on spurious regions. Moreover, we remove the learned spurious correlations with an explanation based learning approach.