SUN_source_code_v2/code/vlfeat-0.9.5/docsrc/aib.html

<!DOCTYPE group PUBLIC "-//W3C//DTD XHTML 1.0 Transitional//EN" "http://www.w3.org/TR/xhtml1/DTD/xhtml1-transitional.dtd">

<group>

<p>The <b>Agglomerative Information Bottleneck (AIB)</b> algorithm
greedily compresses discrete data by iteratively merging the two
elements which cause the mutual information between the data and the
class labels to decreases as little as possible.</p>

<p>Here we test AIB on the problem of finding a discriminatively
optimal quantization of a mixture of Gaussians. The data in this case
is 2 dimensional:</p>

<div class="figure">
<image src="%pathto:root;demo/aib_basic_data.jpg"/>
<div class="caption">
<span class="content">
Random data generated from a Gaussian mixture with three components
(class labels are indicated by color).
</span>
</div>
</div>

<p>We quantize this data on a fixed lattice (a 20x20 grid shown in the
figures below), and construct histograms for each class.</p>

<pre>
f1 = quantize(X1,D,K) ;
f2 = quantize(X2,D,K) ;
f3 = quantize(X3,D,K) ;

Pcx(1,:) = vl_binsum(Pcx(1,:), ones(size(f1)), f1) ;
Pcx(2,:) = vl_binsum(Pcx(2,:), ones(size(f2)), f2) ;
Pcx(3,:) = vl_binsum(Pcx(3,:), ones(size(f3)), f3) ;
</pre>

<p>Next we apply AIB:</p>

<pre>
[parents, cost] = vl_aib(Pcx) ;
</pre>

<p>This provides us with a list of parents of each column
in <code>Pcx</code>, forming a tree of merges. We can now "cut" this
tree to obtain any number of clusters.</p>

<div class="figure">
<image src="%pathto:root;demo/aib_basic_clust_2.jpg"/>
<image src="%pathto:root;demo/aib_basic_clust_3.jpg"/>
<image src="%pathto:root;demo/aib_basic_clust_4.jpg"/>
<div class="caption">
<span class="content">
Three "cuts" of the merge tree, showing 10, 3, and 2 clusters. The
gray squares are nodes of the tree which did not have any data points
which were quantized to them.
</span>
</div>
</div>

<p>Notice that the resulting clusters do not have to be contiguous in
the original space.</p>

</group>