Just Train Twice: Improving Group Robustness without Training Group Information

Standard training via empirical risk minimization (ERM) can produce models\nthat achieve high accuracy on average but low accuracy on certain groups,\nespecially in the presence of spurious correlations between the input and\nlabel. Prior approaches that achieve high worst-group accuracy, like group\ndistributionally robust optimization (group DRO) require expensive group\nannotations for each training point, whereas approaches that do not use such\ngroup annotations typically achieve unsatisfactory worst-group accuracy. In\nthis paper, we propose a simple two-stage approach, JTT, that first trains a\nstandard ERM model for several epochs, and then trains a second model that\nupweights the training examples that the first model misclassified.\nIntuitively, this upweights examples from groups on which standard ERM models\nperform poorly, leading to improved worst-group performance. Averaged over four\nimage classification and natural language processing tasks with spurious\ncorrelations, JTT closes 75% of the gap in worst-group accuracy between\nstandard ERM and group DRO, while only requiring group annotations on a small\nvalidation set in order to tune hyperparameters.\n