Improving Mixup Calibration with Wasserstein Distributionally Robust Optimization
This research introduces DRO-Augment, a training procedure that combines data augmentation with Wasserstein Distributionally Robust Optimization (W-DRO). It regularizes the loss gradient to enhance model resilience against natural data corruptions and adversarial attacks. The method uses a variation regularization approximation of W-DRO to achieve robust generalization bounds.