Thanks for engaging! That makes sense. But I don't in general understand how you decide which parameters are part of p(f) and which are part of p(u). Like you could also have a global u_f, e.g. the topics in a topic model. I understand if you want to want to stop though.