Exact Inference on probabilistic graphical model

Hi everybody,

As I understood, Pyro provides approximate inference such as stochastic variational inference and Monte Carlo based methods. I wonder if there is any plan to develop exact inference such as Variable Elimination?

Thanks you for your time.

Hi @ttc, the only exact inference methods we currently implement are: naive complete enumeration for entirely discrete models, and exact inference in dense Gaussian processes. We do plan to implement message passing for entirely discrete models like HMMs within the next few weeks. Thereafter we may look at Gaussian message passing and generalizations.

Can you point to any particular example models you think would be generally useful?

@fritzo Thanks for replying. I have no particular model in mind at the moment. I just thought that maybe it is educational if we can start from exact inference on some toy example before moving to approximate inference on some large dataset.

A lot of time, I receive some initial datasets from some company and they are actually small so exact inference could be applicable.