Self-Organizing Process Based On Lateral Inhibition And Synaptic Resource Redistribution (1991)
Self-organizing feature maps are usually implemented by abstracting the low-level neural and parallel distributed processes. An external supervisor finds the unit whose weight vector is closest in Euclidian distance to the input vector and determines the neighborhood for weight adaptation. The weights are changed proportional to the Euclidian distance. In a biologically more plausible implementation, similarity is measured by a scalar product, neighborhood is selected through lateral inhibition and weights are changed by redistributing synaptic resources. The resulting self-organizing process is quite similar to the abstract case. However, the process is somewhat hampered by boundary effects and the parameters need to be carefully evolved. It is also necessary to add a redundant dimension to the input vectors.
In Teuvo Kohonen and Kai M{"a}kisara and Olli Simula and Jari Kangas, editors, Proceedings of the 1991 International Conference on Artificial Neural Networks, 415-420, 1991. Amsterdam: North-Holland.

Risto Miikkulainen Faculty risto [at] cs utexas edu