Mixup

Mixup is a data-agnostic augmentation technique that improves neural network generalization by training on convex combinations of input pairs and labels.

Hongyi Zhang and his team at MIT and FAIR introduced Mixup in 2017 to solve the problem of neural network overconfidence. This data-agnostic technique regularizes models by training them on convex combinations of input pairs and their labels. By interpolating features (usually with a Beta distribution and an alpha value of 0.2), the system forces the network to learn smoother decision boundaries. This reduces memorization of noisy data and hardens the model against adversarial attacks. It is a staple in modern computer vision: delivering measurable accuracy boosts on standard benchmarks like ImageNet-2012 and CIFAR-100.

https://arxiv.org/abs/1710.09412