Theoretical Guarantees of Data Augmented Last Layer Retraining Methods

arXiv:2405.05934v1 Announce Type: new
Abstract: Ensuring fair predictions across many distinct subpopulations in the training data can be prohibitive for large models. Recently, simple linear last layer retraining strategies, in combination with data augmentation methods such as upweighting, downsampling and mixup, have been shown to achieve state-of-the-art performance for worst-group accuracy, which quantifies accuracy for the least prevalent subpopulation. For linear last layer retraining and the abovementioned augmentations, we present the optimal worst-group accuracy when modeling the distribution of the …

abstract accuracy art arxiv augmentation combination cs.cv cs.it cs.lg data downsampling fair large models layer linear math.it performance predictions retraining simple state stat.ml strategies training training data type