Mixup

• @zhangMixupEmpiricalRisk2018
• Randomly sample two examples $(x_i,y_i)$ and $(x_j,y_j)$
• New example by weighted 1D piecewise linear interpolation

\hat x = \lambda x_{i}+(1-\lambda)x_{j}

- $$ \hat y = \lambda y_{i}+(1-\lambda)y_{j}

• $\lambda \in Beta Distribution
• New example $(\hat{x},\hat{y})$