event thumbnail image
NIPS '07 Workshop on Approximate Bayesian Inference in Continuous/Hybrid Models
Pascal

Perturbative Corrections to Expectation Consistent Approximate Inference

author: Manfred Opper, TU Berlin

Description

Algorithms for approximate inference usually come without any guarantee for the quality of the approximation. Nevertheless, we often find cases where such algorithms perform extremely well on the computation of posterior moments when compared to time consuming (and in the limit exact) MC simulations or exact enumerations.
A prominent example is the Expectation Propagation (EP) algorithm when applied to Gaussian process classification. Can we understand when and why we can trust the approximate results or, if not, how we could obtain systematic improvements?
In this talk, we rederive the fixed point conditions of EP using the ideas of expectation consistency (EC) [1] and explicitly consider the terms neglected in the approximation. We will show how one can derive a formal (asymptotic) power series expansion for this correction and compute its leading terms. We will illustrate the approach for the case of GP classification and for networks of Ising variables.
[1] Expectation Consistent Approximate Inference, Manfred Opper and Ole Winther, JMLR 6, 2177 - 2204 (2005).

You might be experiencing some problems with Your Video player.
Slides
0:00 Perturbative Corrections for Expectation Propagation
1:31 Outline
2:56 Expectation Propagation (EP) in a nutshell
4:40 Fixed point equations
5:45 The partition function
8:05 EP is optimal to linear order
8:46 Express joint density via q & qn
9:46 Corrections for models with pairwise couplings
11:20 Correction to marginal likelihood (partition function)
13:59 Characteristic function & cumulants
15:35 Express the ratio by cumulants
16:44 Correction to marginal likelihood (partition function)
16:54 Performing the average
17:10 Express the ratio by cumulants
17:34 Performing the average
18:38 Perturbation expansion to Free energy
21:00 Gaussian averages & Feynman graphs - 1
22:01 Gaussian averages & Feynman graphs - 2
23:00 Conjecture: EP is fairly accurate if...
23:37 Gaussian averages & Feynman graphs - 2
24:12 Conjecture: EP is fairly accurate if...
24:21 Gaussian process classification
25:48 The cumulants
25:54 Correction to log partition function
26:48 - Questions
27:26 - Questions
31:04 Correction to log partition function
31:10 Log partition function + correction
31:20 Correcting the posterior mean
31:42 A toy Ising case
32:53 Random Ising networks - 1
33:31 - Questions
34:47 Random Ising networks - 3
35:36 - Questions

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.

Link this page

Would you like to put a link to this lecture on your homepage?
Go ahead! Copy the HTML snippet !

Write your own review or comment: