On the scalability of MSC V ariational inference based on KL(q||p) is scalable in the sense that it works by subsam-4 pling datasets both for exchangeable data, p (x

We thank the reviewers for the constructive feedback, which will significantly improve the paper. We elaborate on this first and address specific comments and questions from the reviewers below. RWS, etc.) applications assumes the data is generated iid and and achieve scalability through use of subsampling and The current discussion in Section 3.5 for MSC on the other hand focuses on the more challenging case, We will clarify this in the revision. We compare the base versions of the respective algorithms. We will add these references to the related work section.