similarity learning
(A shocking example of Andreas Maurers archaic web-design)

Suppose that every time you consult a certain oracle it supplies you with a pair (x,x') of objects together with the information if they are similar or not. The idea of similarity learning is to observe the oracle by repeated consultation and to learn to predict for each pair it produces if its members are similar or not.
I am interested in this subject because it is related to the problem of transfer: Identifying the mechanisms by which the learning of one task can make it easier to learn the next one. A shared concept of similarity can very well be the basis of such a mechanism.
Some of my current work on similarity learning is described in
Learning similarity with operator-valued large-margin classifiers

The paper describes an algorithm for similarity learning and reports some experimental results.
If you are interested in reproducing these results and experimenting with different data-sets you can download a (very rudimentary) executable (Windows XP) from this page, as well as some data-sets which I am referring to in the paper.

You will not be happy with the other files below, if you don't have a Users Manual.
Here is the executable, together with some necessary libraries,
and here are the data-sets for character-recognition under rotation-invariance (alpha, digits), scale-invariance (alpha, digits), and combined rotation and scale invariance (alpha, digits).

Here is also a little visualization program , which shows the algorithm operating on 2-dimensional data.

back to my homepage