![](/files/SOM.svg)
Self-organizing maps (SOMs) are a class of unsupervised learning methods that arrange data samples in a low-dimensional, non-linear manifold in a topology-preserving fashion, i.e. neighbored samples will be mapped to neighbored regions in the manifold. Thus they are well-suited for non-linear dimensionality reduction and visualization. Usually SOMs use discretized maps on a regular grid. Each node on the grid has an associated prototype vector
![$ \mathbf{w}_n $](/files/tex/dd83828edbf154d44d2f29a10af791d5de946f7e.png)
, and learning works in a very similar fashion to vector quantization: pulling the prototype closest to a particular data sample
![$ \mathbf{x} $](/files/tex/1ab3a7df3e68b8e1dedd39f5c4ac0785bd089389.png)
even closer to
![$ \mathbf{x} $](/files/tex/1ab3a7df3e68b8e1dedd39f5c4ac0785bd089389.png)
:
Key to toplogy-preservation is the additional neighborhood function
![$ h_{cn} $](/files/tex/2b3ac723f787fb57a29530140f28d6475f8271a0.png)
, that ensures that not only the winner node
![$ n $](/files/tex/6fcddda10ddd6cd570ff419ee7920043169ccdfd.png)
is adapted, but also its neighbors in the map - weighted exponentially by their distance in the map: