Bounds and estimates for BP convergence on binary undirected graphical models
Description
Belief Propagation (BP) has become a popular method for inference on graphical models. Accurate approximations for intractable quantities (e.g. single-node marginals) can be obtained within rather modest computation times. However, for large interaction strengths (i.e. potentials that are highly dependent on their arguments) or densely connected graphs, BP can fail to converge. This can be remedied by damping the iteration equations, using sequential update schemes or using entirely different algorithms such as double-loop algorithms, that directly minimize the Bethe free energy. However, the value of this approach is questionable, since there is empirical evidence that failure of convergence of BP often indicates low quality of the Bethe approximation.
| Slides | |
| 0:45 | Introduction |
| 3:15 | Graphical model, exact probability distribution |
| 4:53 | Belief Propagation |
| 6:22 | BP for binary variables |
| 7:51 | Norms and contractions |
| 9:07 | Lemma 1. |
| 11:58 | Example: 1-norm for binary variables |
| 13:16 | Example: weighted 1-norm |
| 14:05 | Beyond the binary case |
| 15:38 | We can (try to) bound |
| 18:01 | We can conclude |
| 19:12 | Binary case: comparison of various bounds |
| 29:05 | A very rough average-case analysis |
| 32:43 | Conclusions |
Lecture rating
| People found this lecture: | ||
| Worth seeing | ||
| because it is: | ||
| Valuable and informative | ||
| Well presented | ||
| Easily understandable | ||
| Acceptably recorded | ||
| You need to login to cast your vote. | ||
Report a problem or upload files
If you have found a problem with this lecture or would like to send us extra material, articles, exercises, etc., please use our ticket system to describe your request and upload the data.Enter your e-mail into the 'Cc' field, and we will keep you updated with your request's status.
SEE ALSO:
Link this page
Would you like to put a link to this lecture on your homepage?Go ahead! Copy the HTML snippet !




