Competitive learning using gradient and reinitialization methods for adaptive vector quantization

Abstract

The conventional vector quantization (VQ) for digital coding of large amounts of analog signals, such as image data, speech signals, etc. usually assumes that the input signals follow time-invariant probability distributions. However, the statistics of the signals, sensors, environments, etc. changes slowly in many practical applications. So, we present a competitive learning algorithm for adaptive VQ. We first analyze the gradient method for the competitive learning to adapt to time-varying statistics. To overcome the local minimum problem of the gradient method we present a reinitialization method which embeds the condition of global minimum called equidistortion principle into the competitive learning. By means of computer simulation, we clarify the properties and the effectiveness of the algorithm.