As far as I am reading about VAE, I always see a graphical model of Z --> X,
I know P(X) can be intractable, but that would be if we would have many dependencies. But here is only Z. So why variational inference at first place?:)
$\begingroup$
$\endgroup$
I'm not sure if I get the question correctly, but we use variational inference, because we do have many dependencies. X and Z are usually high-dimensional vectors with complex relationships.
$\begingroup$
$\endgroup$
Exact inference is possible only when your latent variables can take only a discrete set of values. But the computation required grows quite fast with an increase in the dimensionality of latent space. As you change your latent space to a continonus space, the inference becomes intractable.
Hence Inference is the key algorithmic problem. And our goal is to build General and scalable approaches to inference.
