Implementing Neural Gas Networks on Distributed Architectures

Fabio Ancona, Stefano Rovetta, and Rodolfo Zunino

Dept. of Biophysical and Electronic Engineering (DIBE),
University of Genoa, Via all’Opera Pia 11a,16145 Genova, Italy



WCNN96, Proceedings of the 1996 World Congress on Neural Networks,
September 1996. 




Abstract

The paper describes a methodology for the implementation of a
neural algorithm for vector quantization on parallel hardware.
The final application goal is lossy compression of high-dimensional
data for low bit-rate image communication. The high computational
load of the neural training process and the technical importance
of the specific application motivate the search for a highly
efficient parallel implementation of the quantization method.
First, the paper shows a neural algorithm that implement the vector
quantization (Neural Gas). Then, the paper presents the strategy
to distribute the algorithm over an architecture based on a
toroidal mesh topology. Experimental results on an application
testbed consisted in an image-compression task, in which low
bit-rate coding is achieved by Vector-Quantization encoding,
confirm the validity of the parallel approach.