bca82e41ee7b0833588399b1fcd177c7-Reviews.html

The authors propose a parallel algorithm for the DPMM that parallelizes a RJMCMC sampler that jumps between finite models. While the parallelization and the RJMCMC sampler are proposed together, I will separate them for the purpose of this review, in order to ask questions about each part separately. First, the RJMCMC algorithm (by which I mean, the algorithm we would have on a single cluster). Here, we use a reversible-jump MCMC algorithm to jump between finite-dimensional Dirichlet distributions. As an aside, since \bar{\pi}_{K 1} is not used in the mixture model (the mixture model is defined on the renormalized occupied K components), it would seem to make more sense to define a K-dimensional, rather than a K-1 - dimensional, Dirichlet distribution; this is valid under marginalization properties of the Dirichlet distribution, since equation 10 samples from a distribution proportional to \pi_1 ... \pi_K To jump between model dimensionalities, the authors propose a split/merge RJMCMC step that is reminiscent of that of Green and Richardson.