Learning with Neural Methods Computational Intelligence - Learning with Neural Methods on Structured Data

Research of Marc Strickert

Hyperbolic Self Organizing Maps for Sequences (HSOM-S)
Image of Kohonen Map Training Standard SOM is restricted to real number vectors of fixed dimension; thereby sequential information like autocorrelation is lost. This limitation is overcome by adding an explicit back reference to the winner of the previous time step.
 
Adaptation takes into consideration two aspects:
  1. How good is the matching of a neuron to the currently presented sequence element?
  2. Is the candidate neuron also the usual successor of the previous winner?
While part (1) corresponds the ordinary SOM winner neuron search, part (2) takes into consideration the order of input presentation. Question (2) is answered in our HSOM-S approach by explicitly storing and updating the previous winner within each neuron.
 
Context is expressed as a grid position in the triangular grid structure.
 
Spatial or temporal context may be an exponential function of its considered length, therefore an additional feature of HSOM-S is the hyperbolic neuron neighborhood connectivity.

LNM - Home: Computer Science - University of Osnabrück  | 27 September 2003