Globally-ordered topology-preserving maps achieved with a learning rule performing local weight updates only

Abstract

A new unsupervised competitive learning rule is introduced for topology-preserving map formation and vector quantization. The rule, called maximum entropy learning rule (MER), achieves a globally-ordered map by performing local weight updates only. Hence, contrary to Kohonen's self-organizing map algorithm and its many variations, no neighborhood function is needed. The rule yields an equiprobable quantization of a d-dimensional input p.d.f. Simulations are performed to show that the dynamical- and convergence properties of MER are essentially different from those of Kohonen's algorithm.