mixup

Mixup is a data-augmentation and regularisation method that trains the network on linear blends of pairs of examples. For two training images and their one-hot labels, it forms a new training example by taking a convex combination: the mixed image is lambda times the first image plus (1 minus lambda) times the second, and the mixed label is the same weighted average of the two label vectors. The mixing weight lambda is drawn randomly from a Beta distribution for each pair.

Because both the pixels and the targets are interpolated, the model is asked to predict, say, 70 percent cat and 30 percent dog for an image that is a 70/30 pixel blend. This encourages the network to behave linearly between training examples, which empirically smooths decision boundaries, reduces memorisation of noisy labels, improves generalisation and calibration, and increases robustness to adversarial perturbations.

Mixup is cheap, architecture-agnostic, and combines well with standard augmentations, so it is common in strong classification recipes. Its main subtlety is that the blended images can look unnatural, and a too-aggressive Beta parameter, meaning heavy mixing, can underfit; practitioners tune the Beta concentration, often around 0.2, and frequently pair mixup with its spatial cousin CutMix.