The content discusses the issue of spurious correlations in machine learning models, where standard empirical risk minimization (ERM) models tend to prioritize learning spurious correlations between spurious features and true labels, leading to poor accuracy on groups where these correlations do not hold.
To address this problem, the authors propose GIC (Group Inference via data Comparison), a novel method that accurately infers group labels by leveraging a comparison dataset with a slightly different group distribution. GIC trains a spurious attribute classifier based on two key properties of spurious correlations: (1) high correlation between spurious attributes and true labels, and (2) variability in this correlation between datasets with different group distributions.
The authors demonstrate that the inferred groups from GIC can be seamlessly integrated with various downstream invariant learning algorithms, such as Mixup, GroupDRO, Upsample, and Subsample, to improve the worst-group accuracy. Empirical studies on multiple datasets show that GIC consistently outperforms existing group inference methods in terms of recall and precision, and can even match the performance of methods that use oracle group labels.
Furthermore, the authors analyze the misclassifications in GIC and identify an interesting phenomenon called "semantic consistency", where GIC tends to assign similar semantic instances to the same group, even if they are not categorized into the same group by human decisions. This semantic consistency can benefit methods like Mixup, which rely on distorting semantics for invariant learning, leading to improved worst-group accuracy compared to using oracle group labels.
To Another Language
from source content
arxiv.org
Principais Insights Extraídos De
by Yujin Han,Di... às arxiv.org 04-23-2024
https://arxiv.org/pdf/2404.13815.pdfPerguntas Mais Profundas