Self Organizing Map (SOM)

Self Organizing Map (SOM)

In 2014, the need for projection from multi-dimensionnal spaces toward 2D or 3D appears. From the 12 audio indicators provided by Echonest to 2 dimensions or 1 dimension. We had already used PCA (principal Component Analysis) but SOMs and their algorithms were facinating.

Therefore, I coded a generic java package.

The java class needs only a distance function over a set of points. The topology is set only through the distance function. The change of learning factors over iterations are specified using a linear fonction and key points.

The following illustrations show various kinds of use.

A video shows the package in use, there.

A SOM on plane with toric neighbours
A 1D SOM
A spherical SOM
Another sherical SOM
Location of music styles on a sphere