Aligned SOMs

The pieces of music are grouped using a neural network algorithm, the Self-Organizing Map (SOM). The SOM projects high-dimensional data (describing, for example, rhythm or timbre) to a 2-dimensional plane. Songs with similar high-dimensional representations are located close together on the map.

A Short Introduction to SOM-Training

The SOM consists of a number of map units. Each of these map units is assigned a model vector with the same dimensionality as the data items. The initial values of the model vectors can be assigned randomly or on the basis of the data set. Training is done by presenting the data items to the SOM either one after another (sequential training) or all at the same time (batch training). For each data item, the map unit with the most similar model vector (called best matching unit or BMU) is calculated according to a distance metric (e.g. Euclidean distance). Hereafter, this unit as well as its neighbors are updated. That means the values of the model vectors are changed according to the data item. Map units close to the best matching unit are modified more significantly than those farther away. This process is repeated until the SOM converges, i.e. the model vectors do not change remarkably any longer.

Aligned SOMs

Aligned Self-Organizing Maps allow for changing the view according to different weightings of the underlying data. The example below shows 5 aligned SOMs with gradually decreasing influence of rhythm and increasing influence of timbre. To enlarge, simply click on the respective image.