Implementation and Evaluation of Large Scale SOM in GPU Cluster

Abstract

When the SOM learning algorithm applies to a large amount of data, it can be an effective approach that a parallel and distributed processing is adopted. There are two approaches in parallelization of SOM learning algorithms. One is by dividing the learning dataset and another is by dividing a SOM’s competitive layer. Furthermore, GPGPU approach is also adopted for vector operations under SOM learning process and we have confirmed an effectiveness of these implementation methods with respect to a computation time. In this paper, we present a way of implementation of each two kind of parallelized SOM and its evaluation.