|
Hi, I am having difficulty getting my head around variational methods for bayesian inference. I understand Monte Carlo methods for approximate inference in bayesian networks but seem to have difficulty understanding variational methods. Is there any accesible resource for understanding that. And how does one go about developing variational inference methods ? |
|
A variational method is any method which replaces the integration step necessary to compute the normalizing constant of a probability distribution with an optimization problem. There are a few known ways of doing so, each of which makes slightly different assumptions about the model and works in slightly different ways. A simple google search will point you to many good tutorials on variational methods, for example this one by david blei or this one by jason eisner. I also like the explanation in Matt Beal's thesis. |
