English (United Kingdom)

https://curated-unify.zendy.io/wp-json/zendy-region/v1/featured_content/oa?rat=en

https://curated-unify.zendy.io/wp-json/zendy-region/v1/highlighted_journal/

Global: Zendy Plus annual

Presents the access of premium content as premium feature

Premium Content

Presents the keyphrase highlighting as premium feature

Keyphrase Highlighting

Presents the summarisation as premium feature

Summarisation

Insights

Presents the pdf analysis as premium feature

PDF Analysis

Presents the zaia usage as premium feature

ZAIA

Global: Zendy Plus monthly

Global: Zendy Tools annual

Global: Zendy Tools monthly

Global: Zendy Free Plan

Deep neural networks trained by minimizing the average risk can achievestrong average performance. Still, their performance for a subgroup may degradeif the subgroup is underrepresented in the overall data population. Groupdistributionally robust optimization (Sagawa et al., 2020a), or group DRO inshort, is a widely used baseline for learning models with strong worst-groupperformance. We note that this method requires group labels for every exampleat training time and can overfit to small groups, requiring strongregularization. Given a limited amount of group labels at training time, JustTrain Twice (Liu et al., 2021), or JTT in short, is a two-stage method thatinfers a pseudo group label for every unlabeled example first, then appliesgroup DRO based on the inferred group labels. The inference process is alsosensitive to overfitting, sometimes involving additional hyperparameters. Thispaper designs a simple method based on the idea of classifier retraining onindependent splits of the training data. We find that using a novelsample-splitting procedure achieves robust worst-group performance in thefine-tuning step. When evaluated on benchmark image and text classificationtasks, our approach consistently performs favorably to group DRO, JTT, andother strong baselines when either group labels are available during trainingor are only given in validation sets. Importantly, our method only relies on asingle hyperparameter, which adjusts the fraction of labels used for trainingfeature extractors vs. training classification layers. We justify the rationaleof our splitting scheme with a generalization-bound analysis of the worst-grouploss.

Improved Group Robustness via Classifier Retraining on Independent Splits