pyro-ppl / funsor

Functional tensors for probabilistic programming
https://funsor.pyro.ai
Apache License 2.0
236 stars 20 forks source link

Implement unbiased wrappers for biased interpretations #127

Open fritzo opened 5 years ago

fritzo commented 5 years ago

We should be able to wrap a biased approximate interpretation in an unbiased interpretation using rejection sampling with a biased proposal. For example we should be able to use moment_matching #115 , linearize #89 , and EP #90 as proposal distributions.

Kevin Murphy points out:

I think these fast, deterministic locally approximating methods [like moment matching and Gaussian approximation] could be really useful as proposals for MC methods. Here is the paper by Nando that I mentioned (I am sure there are many more papers like this :)

R. van der Merwe, A. Doucet, N. de Freitas, and E. A. Wan, “The Unscented Particle Filter,” in Advances in Neural Information Processing Systems 13, T. K. Leen, T. G. Dietterich, and V. Tresp, Eds. MIT Press, 2001, pp. 584–590 [Online]. Available: http://papers.nips.cc/paper/1818-the-unscented-particle-filter.pdf

fritzo commented 5 years ago

@eb8680 @lawmurray @martinjankowiak any ideas how to best do this in Funsor?

eb8680 commented 5 years ago

Another interesting paper on a similar theme: http://papers.nips.cc/paper/8041-graphical-model-inference-sequential-monte-carlo-meets-deterministic-approximations