Reviews: MetaReg: Towards Domain Generalization using Meta-Regularization

MetaReg in a nutshell: This paper reinterprets and further develops few-shot meta-learning ideas for the challenging domain generalization paradigm, using standard supervised benchmarks. The main contribution is the learning of a regularizer, as opposed to learning an initial set of parameters well "positioned" for finetuning. Scores seem to be significantly improved in several cases, but I am not an expert. Pros: - The paper goes beyond the original inspiration and adapts the approach to serve a substantially different problem. While in meta-learning the degree of similarity between problem instances is substantial, e.g.