Criticism of VAE


#1

Hi @dustin,
Could you provide an example about the way too criticise VAE.

I tried the example in Edward example folder, but creating y_post and using ed.evaluate() raised an error.


#2

The graphical model for a VAE has fixed model parameters and a set of local variables z per data point x. This means that to generate new data points, you need to sample from the prior, then sample from the likelihood (which is what we do in the VAE examples). There is no posterior predictive to build.