Variational Inference for Bayesian Networks Logo
[Sourceforge project page]
John Winn, January 2004
Overview Tutorial Examples Help

Online Help: Using VIBES from Matlab

This is BETA documentation

Installing VIBES into Matlab

Download the VIBES Jar file (see the Overview page) and place it in [matlab directory]/vibes. Then edit [matlab directory]/toolbox/local/classpath.txt to add the following lines

## VIBES classes
[matlab directory]/vibes/Vibes2_0.jar

To test the installation, run Matlab and type

>> import*;

This imports VIBES functionaility into Matlab and needs to be done once per session. Now type,


This should cause VIBES to run normally from within Matlab. Of course, you will want to use VIBES programmatically instead, so read on!


In the beta, VIBES exposes an API which allows pre-defined networks to be loaded and inference to be performed. The following example outlines the API. If you want to work through it yourself, you need to download the Gaussian network file.

Loading and displaying a network

To load a VIBES network into Matlab, use:

>> net ='Gaussian.xml')

which will give a result like

net =
ModelNetwork Demo1 - Gaussian [3 nodes]
  ModelNode x [0 children] in (400x64)
  ModelNode beta [1 child] in (64)
  ModelNode mu [1 child] in (64)

For debugging purposes, you can display the loaded network with


Setting observed values and performing inference

You can dynamically changes the size of plates in your network, according to the dimensionality of your data, for example

>> N = 50; d= 3;
>> net.getPlate('N').setSize(N)
>> net.getPlate('d').setSize(d)

Now you can attach observed data

>> xdata = [randn(N,1) randn(N,1)+1 randn(N,1)+2];
>> net.getNode('x').setData(xdata)

To perform inference, you use an inference algorithm object which may be obtained by initialising the network

>> algorithm = Vibes.init(net);

A number of iterations of variational inference (in this case 50) can be performed using

>> algorithm.update(50);

The indicated number is a maximum. If inference converges, the algorithm will stop earlier.

If desired, the default update order can be overridden by calling

>> Vibes.setUpdateOrder(algorithm,'mu,beta');

prior to calling update.

Extracting or initialising the variational distribution moments

Following initialisation or inference, you can extract the moments of the variational distribution for a given node using Vibes.get(), for example

>> mu=Vibes.get(net,'mu');
mu =

    0.1083    0.0270
    1.0060    1.0273
    1.9381    3.7764

The returned matrix has size d2. As the mu node is contained in the d plate, there are d copies of it and the rows of this matrix contain the moments of the variational distribution for each copy. As mu is Gaussian, the first column gives <mu> and the second gives <mumu>.  The supplied moments will be different for other distributions.

Hence, to get the variance for each copy of mu, type

>> mu(:,2) - mu(:,1).^2

Prior to performing inference (but after initialisation), you can directly modify the initial variational distribution of a node by setting its moments using

>> Vibes.set(net,'mu', initial_mu);

The matrix of moments supplied must have the same form as that returned by Vibes.get().