Derivation of ELBO upon the Existence of Conditional Latent Variable Model

93 Views Asked by At

I am reading the recently published paper from DeepMind, "Neural Scene Representation and Rendering" and especially its "Supplementary Materials".

Following is the page 1 and it's pretty hard for me to derive the upper-bound $F(\theta, \phi)$ equal to $-L(\theta) + \sum KL_{divergence}$

I think there might be more implicit steps to get to the conclusion but can't find something nicely fit deriviation passage.

Any help or advice would be appreciated.

enter image description here