Self-Organizing Maps (SOM)
Self-Organizing Maps (SOM: Kohonen, 2001) are neural networks that adapt to regularities in input data using unsupervised learning. Representations learned with SOM have interesting topological properties, namely the fact that input patterns that are similar are coded on regions of the SOM that are nearby.
The Matlab implementation features receptive fields, and presents two applications related to visual word recognition:
- Learning letter representations from retinal images (run main.m in ./code/RetinalImageToLetters)
- Learning sub-lexical representations from letters (run main.m in ./code/LettersToSubLexicalUnits)
Download (version 1.1.1, Matlab code, 301 kb ZIP file)
- Kohonen, T. (2001). Self-Organizing Maps (Third, extended edition). Springer.
- See also: http://en.wikipedia.org/wiki/Self-organizing_map