Curriculum-enhanced GroupDRO: Challenging the Norm of Avoiding Curriculum Learning in Subpopulation Shift Setups

Researchers propose Curriculum-enhanced Group Distributionally Robust Optimization (CeGDRO), a new machine learning approach that challenges conventional wisdom by using curriculum learning in subpopulation shift scenarios. The method achieves up to 6.2% improvement over state-of-the-art results on benchmark datasets like Waterbirds by strategically prioritizing hard bias-confirming and easy bias-conflicting samples.

• →CeGDRO introduces curriculum learning to subpopulation shift problems, breaking from traditional approaches that avoid this technique.
• →The method initializes model weights at an unbiased position to prevent convergence toward biased hypotheses.
• →Results show consistent improvements across all tested subpopulation shift datasets with up to 6.2% gains on Waterbirds.
• →The approach prioritizes hardest bias-confirming samples and easiest bias-conflicting samples during training.
• →This research demonstrates that curriculum learning can be beneficial in scenarios previously thought unsuitable for such methods.