0
$\begingroup$

For example, if I define my prior distribution as Gaussian but I define my likelihood as multinomial, does this break any theoretical basis for variational inference?

$\endgroup$
1
$\begingroup$

No this breaks no theoretical basis for Bayesian inference, nor for Variational inference. However, it may make your computation more complicated. Depending on the form of variational inference this may be more or less of a concern. For example, in mean-field variational inference you will likely break the nice, closed form solutions you get when you choose the conjugate priors.

$\endgroup$
  • $\begingroup$ thank you! are you familiar with the Variational autoencoder or stochastic gradient Variational bayes? if so, do you know if it matters if one uses a multinomial likelihood with a gaussian prior? $\endgroup$ – wcarvalho Aug 16 '17 at 18:52
  • $\begingroup$ I'm not familiar w/autoencoders, SGD is an optimization method and VB is a form of variational inference. Neither of these two preclude the use a multinomial or a Gaussian prior. $\endgroup$ – Lucas Roberts Aug 16 '17 at 18:59

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.