An efficient algorithm for parallel distributed unsupervised learning

This paper presents a new technique for parallel and distributed unsupervised learning. It arises from a detailed analysis of the weaknesses of several, existing algorithms, the most important of which is the presence of intrinsically serial operations. The basic idea of this work, therefore, is the substitution of this latter with new operations, as similar as possible to the original ones, but better suited to a parallel implementation. The result is a notable increase in speed-up; the price to be paid is a slight deterioration in the precision of the clustering process. The ideal applications for the new algorithm are very complex problems with a high number of patterns and classes.