Conditional vs. nonconditional variational family

Question

In variational bayes, distributions in the variational family $\mathcal{Q}$ are denoted $q_\phi$ and are used to approximate the posterior $p_\theta(z|x)$. However, I've seen both notations $q_\phi(z|x)$ and $q_\phi(z)$. What's the difference and how are they each used?

My interpretation was that $q_\phi(z)$ and $q_\phi(z|x)$ are a prior and posterior on $\mathcal{Q}$. Is this the right interpretation? It doesn't make sense that we should need a prior for $q_\phi$ since it is being optimized as an approximation to $p_\theta(z|x)$.

Answer 1

"My interpretation was that $q_{\phi}(𝑧)$ and $q_{\phi}\left(z | x \right)$ are a prior and posterior on $\mathcal{Q}$" is not correct at all. $\mathcal{Q}$ is the variational family defining the distribution space. $q_{\phi}(𝑧)$ and $q_{\phi}\left(z | x \right)$ live in $\mathcal{Q}$ and used to approximate the prior/posterior defined for a data distribution.

Both notations are used actually. $q_{\phi}(𝑧)$ actually stems from the ELBO derivation(You can check these lecture notes from CMU), i.e. the KL divergence definition.

For VAE's, we specifically denote $q_{\phi}\left(z | x \right)$ because the encoder tries to learn an amortized posterior distribution, which instead of optimizing a set of free parameters unlike in the lecture notes that I've shared sbove, we can propose parameterized function(like a neural network) that maps from observation space to the parameters of the approximate posterior distribution.

In fact, there are advanced VI techniques that allow implicit priors for VAE's like Semi-Implicit Variational Inference (SIVI).

TL;DR, try to stick with $q_{\phi}\left(z | x \right)$ for VAE posteriors and always make sure to understand which distribution you are trying to approximate.

Hope this helps.

Answer 2

Note: the point of Amortized VI is this:

Traditional VI defines the latent variable $z_i$ for each individual observation, and then optimizes them all jointly. This is computationally expensive to have to re-run the inference procedure for every datapoint. Instead, one can define a parameterized function, i.e. $q_\phi(z|x)$ sending $x \rightarrow z$. Now the parameters to update are the parameters of the function rather than the distributions of every $z_i$. For a new observation, we only have to pass it through the network rather than update everything.