Label smoothing is a very effective technique in which we are asking our model to predict some value smaller than 1 for the correct label. This has a regularizing effect and also helps greatly with training on noisy labels. It is a great technique to use with mixup.