1
$\begingroup$

Typically, we use Gibbs sampling to update (or generate samples from) energy based models. This means we update each node while keeping its markov blanket constant.

Why can't we update/sample all the nodes at the same time?

If your answer is “because otherwise we can't reach thermal equilibrium”, then why can't we reach thermal equilibrium?

$\endgroup$
3
  • $\begingroup$ To rephrase my question: What would go wrong (and why), if we used the conditional distributions to update all nodes at once? $\endgroup$
    – Sia Rezaei
    Commented Jan 6, 2020 at 16:51
  • $\begingroup$ Sure, but how would this update scheme differ from Gibbs update scheme in practice? For example, would this suffer from some kind of bias in the samples it generates, compared to Gibbs sampling? and why? $\endgroup$
    – Sia Rezaei
    Commented Jan 9, 2020 at 16:35
  • $\begingroup$ Yes, I know what Gibbs sampling is. My question is how would the distribution of the generated samples be different between these two updating schemes. $\endgroup$
    – Sia Rezaei
    Commented Jan 10, 2020 at 17:26

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Browse other questions tagged or ask your own question.