Mixup Without Hesitation

Mixup linearly interpolates pairs of examples to form new samples, which has been shown be effective in image classification tasks. However, there are two drawbacks mixup: one is that more training epochs needed obtain a well-trained model; the other mixup requires tuning hyper-parameter gain appropriate capacity. In this paper, we find constantly explores representation space, and inspired by exploration-exploitation dilemma, propose Without hesitation (mWh), concise algorithm. We show mWh strikes good balance between exploration exploitation gradually replacing with basic data augmentation. It can achieve strong baseline less time than original without searching for optimal hyper-parameter, i.e., acts as hesitation.