Learning-AI

mixup: Beyond Empirical Risk Minimization

Mar 2019

tl;dr: Linear blending of labels boosts accuracy and generalizability of classification.

Overall impression

The mixup technical is simple as it does not require domain knowledge (data agnostic), but it is surprisingly effective. It can be seen as a special form of data augmentation, and also as a regularization technique.

Key ideas

Technical details

Notes