ELBOs defined by an SMC's marginal likelihood estimation

Misc.
1

Hello!

Have there been any past efforts, or are there any future plans, to implement an ELBO loss based on marginal likelihood estimation using a particle filter (SMC)? Relevant papers: Filtering Variational Objectives, Variational Sequential Monte Carlo, Auto-Encoding Sequential Monte Carlo.

I’m interested in discussing the possibility of implementing these methods, as I am currently experimenting with differentiable resampling algorithms for particle filters and would like to use Pyro for implementation. Additionally, I would appreciate any guidance from a core developer.

2

i am not aware of any such plans past or present.

it would of course be possible to do so but i suppose for practical reasons one would probably want to commit to some relatively restricted class of models (no nested time series or any such craziness)