Adaptive vector quantization with a structural level adaptable neural network

Abstract

A new type of adaptive vector quantizer is proposed. The core of this system is a self-development neural network constituting the codebook of the vector quantizer. Each neuron in the network memorizes a codeword of the active codebook in its input interconnection weight vector. The codebook is constantly evolving with time to reflect the statistical fluctuation of the source signals. The dynamics of the codebook is characterized by neuron generation, neuron weight vector adjustment, and neuron annihilation processes of the network. The quantization residue of the neural network quantizer is fed to a fixed structure lattice vector quantizer, and the quantized residue is used to stimulate the evolution process of the neural network codebooks inside both the transmitter and the receiver.<>