PAUL: a Parallel Algorithm for Unsupervised Learning

G. Campobello, G. Patanè and M. Russo 

Abstract

In this paper the Parallel Algorithm for Unsupervised Learning (PAUL) is presented. Its ideal field of application is constituted by very complex problems with a high number of patterns and classes to be considered. In such cases, PAUL benefits from the use of parallel or distributed resources as regards both the computational power and the availability of physical memory. The fundamental idea from which this work arises is to perform the parallelization of an algorithm by substituting intrinsically serial operations with efficiently parallelizzable versions that try to approximate the original ones as well as possible. The results obtained are a slight deterioration of the final quantization error and a sensible improvement of the speed up with respect to the value we would get by parallelly implementing the original algorithm leaving all of the computational load to a single task.

Keywords: Clustering, Vector Quantization, Unsupervised Learning,LBG, ELBG, PAUL

[MyPaul] G. Campobello, G. Patanè, and M. Russo, PAUL: a Parallel Algorithm for Unsupervised Learning. IEEE Transactions on Neural Networks, submitted