<ul><li>The authors propose a class-specific scaling strategy for group distributionally robust optimization to improve robust accuracies and mitigate spurious correlations and dataset bias.</li><li>They further develop an instance-wise adaptive scaling technique to alleviate the trade-off between robust and average accuracies.</li><li>A na"ive ERM baseline using the proposed class-specific scaling technique matches or even outperforms recent debiasing methods.</li><li>They also introduce a novel unified metric to quantify the trade-off between robust and average accuracies.</li></ul>

Re-evaluating Group Robustness via Adaptive Class-Specific Scaling

Discover more