A competitive and selective learning method for designing optimal vector quantizers
N. Ueda, Ryoko Nakano · 2002
A new competitive learning method with a 'selection' mechanism is proposed for the design of optimal vector quantizers. A basic principle called the 'equidistortion principle' for designing optimal quantizers is derived theoretically, and a new learning algorithm based on this principle is presented. Unlike conventional algorithms based on the 'conscience' mechanism, the proposed algorithm can minimize distortion without a particular initialization procedure, even when the input data cluster in a number of regions in the input vector space. The performance of this method is compared with that of the conscience learning method.>