Export Reviews, Discussions, Author Feedback and Meta-Reviews

Review Summary Score: 6, Marginally above the acceptance threshold The proposed method for streaming, distributed inference of DP mixture models presents a nice solution to the cluster identification problem, backed by experiments that are convincing though not rock solid. I'm hesitant to recommend unconditional acceptance, because basic information about how new clusters are created at each minibatch are totally absent, hurting reproducibility. Summary of Paper This paper develops a new algorithm for streaming, distributed variational inference for the DP mixture model, with some supplementary material suggesting how to use these insights for many other BNP models. Using a mean-field approximation, the authors consider how to allow multiple worker nodes to process data batches in parallel and then aggregate these results asynchronously. In particular, the authors offer a new solution to the "component identification" problem: how to find correspondence between new clusters created independently by two separate worker nodes.