A counter-intuitive augmentation where two images and their labels are linearly blended (e.g., 60% Dog + 40% Cat). It encourages the model to behave linearly in-between training examples.
Regularization technique.
improves robustness and generalization.