A self-organizing map (SOM) or self-organizing feature map (SOFM) is a type of artificial neural network (ANN) that is trained using unsupervised learning to produce a low-dimensional (typically two-dimensional), discretized representation of the input space, called a map. Self-organizing maps are different from other artificial neural networks in the sense that they use a neighborhood function to preserve the topological properties of the input space. Read more...

On radio aporee, we're researching the possibility to create non-geographical topographic maps out of the sounds and their meta data. Since it is impossible to create an exact 2-D image of n-dimensional data, we rather try to find tendencies, proximity and clusters of similarity. Distance implies difference, but nearness does not necessarily imply resemblance. Lines and structures between areas indicate divergence, like rivers and mountains separate nearby areas in real landscapes. How meaningful these maps can be depends on the pattern we choose for comparision. Data analysis is based on Peter Kleiweg's implementation of Teuvo Kohonen's algorithm.

Currently implemented on aporee:

• selection (sounds < 10min.) of the radio aporee :: maps sounds, compared by audio data:
  • min/max/mean/rms amplitudes
  • min/max/mean/rms deltas of consecutive audio samples
  • average frequency
  • peak count (how often does the audio peak max/min amplitude)
  • crest factor (standard ratio of peak to RMS levels)
  • flat factor (consecutive samples with the same value)
  • map size: 100x100, learning epoches: 200
• same data as above, map size: 75x75, 100 learning epoches
• same data as above, map size: 50x50, 50 learning epoches