Sharpness-aware minimization (SAM) has been shown to improve the generalization of
neural networks. However, each SAM update requires\emph {sequentially} computing two …