mixup

This is a method for making neural networks more robust by creating virtual training examples through convex combinations of input features and their corresponding labels.