[dgmdoug]

I think it's more to do with the application of parallel MC to approximate inference over Bayesian networks. In the normal case of approximate inference, each node would operate over the incoming data to approximate maximum likelihoods (this is the MC bit) then pass them out on each link. The key contribution of this work seems to be to tie the MC algorithms together from difference nodes. Difficult to say any more without reading the paper though.

My key question would be whether there are independence considerations that Parallel MC breaks? (Message passing works because all nodes need only information from their neighbours to summarise the rest of the network, due to the modelled independence).