<ul><li>Deep learning models often learn spurious correlations between targets and non-essential features, leading to spurious bias that hampers model performance on data lacking these correlations.</li><li>Existing methods for mitigating spurious bias require group labels that involve costly human annotations, which may not capture subtle biases like relying on specific pixels for predictions.</li><li>A new framework called ShortcutProbe is proposed, which does not rely on group labels to mitigate spurious bias. It identifies prediction shortcuts in a model's latent space and retrains the model for improved robustness.</li><li>The ShortcutProbe framework is shown to be theoretically effective and practically efficient in enhancing a model's robustness to spurious bias across various datasets.</li></ul>

ShortcutProbe: Probing Prediction Shortcuts for Learning Robust Models

Discover more