This paper presents a modular meta-learning method where Bayesian shrinkage is used to decide which modules to be updated for a new task. All reviewers agree that the modular approach has a clear motivation, is theoretically sound, and is thoroughly validated by a large set of experiments.