Review for NeurIPS paper: MetaPerturb: Transferable Regularizer for Heterogeneous Tasks and Architectures

The idea of meta-learning perturbation function is not new. As the authors mentioned, [A] also meta-learns a kind of perturbation. This work seems an incremental version of the previous paper such as [A]. For example, Eq. (1) in this paper is exactly the same as Eq. My understanding is that this new paper applies a better noise generator than that of [A].