Catalogue of Artificial Intelligence Techniques
Self-organising Feature Maps
Keywords: neural networks
Categories: Data Models , Knowledge Representation , Pattern Recognition and Image Processing , Speech , Neural Networks
Author(s): Steve Renals
Self-organising feature maps were introduced by Kohonen in 1982. They perform a dimensionality reduction, with a multi-dimensional input being projected onto a two-dimensional feature map. A key notion here is that of conservation of topology--points that are close together in input space are mapped in such a way that they remain close in the two-dimensional feature space. This notion of dimensionality reduction is obviously inapplicable to uniformly distributed input data; however it transpires that many real-world problems may be mapped in such a way without a catastrophic loss of information. An example of the use of this algorithm is in speech recognition, whereby high-dimensional speech data is mapped onto a 2-dimensional `phonotopic map'.
- Kohonen, T., Self-organisation and associative memory
, Springer-Verlag, New York , 1990 (3rd edition