I used growing neural gas ( http://web.cs.swarthmore.edu/~meeden/DevelopmentalRobotics/f... ) for my MS research dealing with keystroke dynamics (specifically, modeling Hidden Markov Models on interkey-delays to identify a user's language). When you have a large set of data and you need to identify exemplars for that data, it does a pretty good job. As _delirium said, though, it requires a degree of supervision and tweaking to get growth parameters, etc.

Edit: I use 'supervision' in a 'it might take a few tries to get it working ideally' sense, not a 'you need a priori knowledge of the data' sense.