Argonauts:Self-organizing maps
From Wasteland
Contents |
18 February 2006
Spatio-temporal coordinates
WMVL, 11.00am to 1.00pm (usually exactly, but not this time!).
Attendees (in alphabetical order)
- Ramón Casero Cañas.
- Mike Kadour.
- Etienne von Lavante.
- Rohan Loveland.
Minutes
Niranjan suggested Ch. 9 of the book
S. Haykin. Neural Networks. A Comprehensive Foundation. 2nd Ed. Pearson Education. 1999.
It took us 15 min to make photocopies for everyone. This is why. To photocopy a book, you need to measure it, and select the option "Custom Size" in the evil photocopier. Then you select A4 size, but horizontal, and the option to convert from one-sided to two-sided. And of course, if you don't select the "Separate" option, all copies come interlaced and then you have to go one by one to put them in 4 heaps. Now, this sounds easy, but don't forget that we are a bunch of Engineers. Our promise: "No job too easy, no mistake too big".
Anyway all this was in vain, because after handing the photocopies somebody said: "Aw, this book looks too difficult, let's google for something simpler instead".
Self-Organizing Maps (SOMs) are neural networks (NNs) for unsupervised learning, usually with a 2D lattice of neurons. At the input we have samples of high-dimensional data, for example X = (age, weight, height, blood pressure, heart rate, ejection fraction). At the output, the data is clustered on the NN lattice.
In the learning phase, each sample gets a maximum response from one of the neurons. The connectivity of this neuron and of its neighbourhood is reinforced, and the connectivity of the rest of the neurons is reduced. Both the size of the local neighbourhood and the learning rate are dynamically adjusted in each iteration of the algorithm.
Cool, isn't it? To the point that some people tried an implementation of the whole thing, couldn't finish before 1pm and continued working on it for another 20 min. This is most certainly to be considered heresy.
--
