Suppose you want to find a way to cluster a bunch of objects (e.g. genes
or images).
Take k random subsets of the objects of size s each.
Assign randomly s/2 a label of +1 and s/2 a label of -1.
Find the separating hyperplane using SVM or something.
Now characterize every object based on their values for the k
resulting hyperplanes.
Two objects are close if they have similar values in the k hyperplanes.
Awesomely cool and simple.
http://deuxalex.free.fr/rmmh-cvpr-certified-IEEE-eXpress.pdf