Multilevel Parallelization of Unsupervised Learning Algorithms in Pattern Recognition on a Roadrunner Architecture

摘要

The aim of the paper is to present a solution to the NP hard problem of determining a partition of equivalence classes for a finite set of patterns. The system must learn the classification of the weighted patterns without any information about the number of pattern classes, based on a finite set of patterns in a metric pattern space. Because a metric is not suitable in all the cases to build an equivalence relation, an ultrametric is generated from indexed hierarchies. The contributions presented in this paper consists in the proposal of multilevel parallel algorithms for bottom-up hierarchical clustering, and hence for generating ultra-metrics based on the metrics provided by the user. The algorithms were synthesized and optimized for clusters having the Roadrunner architecture (the first supercomputer that breaks lPFlops barrier [1]).