Freeze then Train: Towards Provable Representation Learning under Spurious Correlations and Feature Noise

Open in new window