Improving Discriminative Visual Representation Learning via Automatic Mixup

Mixup, a convex interpolation technique for data augmentation, has achieved great success in deep neural networks. However, the community usually confines it to supervised scenarios or applies it as a predefined augmentation strategy in various fields, grossly underestimating its capacity for modeling relationships between two classes or instances. In this paper, we decompose mixup into two sub-tasks of mixup generation and classification and formulate it for discriminative representations as class- and instance-level mixup. We first analyze and summarize the properties of instance-level mixup as local smoothness and global discrimination. Then, we improve mixup generation with these properties from two aspects: we enhance modeling non-linear mixup relationships between two samples and discuss learning objectives for mixup generation. Eventually, we propose a general mixup training method called AMix to improve discriminative representations on various scenarios. Extensive experiments on supervised and self-supervised scenarios show that AMix consistently outperforms leading methods by a large margin.