Self-development competitive learning VQ based on vitality conservation networks

Abstract

A novel self-development network effective in competitive learning vector quantization, called PVC (periodical vitality conservation) is proposed. Each neuron is associated with a value of vitality, a measure of winning frequency during the successive input adaptation process. Conservation is achieved by keeping the total sum of vitality at constant 1, as vitality values of all neurons are updated after each input presentation. Conservation in vitality facilitates systematic derivations of learning parameters, including the learning rate control which greatly affects the performance. Extensive comparisons of PVC and other self-development models are also presented. Simulation results show that PVC is very effective in learning a near-optimal vector quantization in that it manages to keep a balance between the equi-probable and equi-error criteria.