EVOLUTION OF INTERNAL REPRESENTATIONS GENERATED BY UNSUPERVISED SELF-REFERENTIAL NETWORKS

A novel learning algorithm for unsupervised topological clustering is introduced that generates mappings of arbitrary randomly received inpu t signals. The concept of a self-referential adaptation is defined; th is leads to a dynamical assignment of equivalence classes. The optimiz ed topological representation allows to distinguish between even very similar input vectors. The performance of the algorithm is analysed st atistically and conclusions for the mathematical description of self-r eferential biological systems are derived.