Parallel Batch Self-Organizing Map on Graphics Processing Unit Using CUDA
- 561 Downloads
Batch Self-Organizing Map (Batch-SOM) is being successfully used for clustering and visualization of high-dimensional datasets in a wide variety of domains. Although the structure of its training algorithm has a high potential for parallelization, focus of the previous efforts has been on the original Step-wise SOM. This gap is due to the facts that Batch-SOM requires some extra precautions (specially in its initialization phase), and it took quite a while since its introduction that researchers affirmed the desirability of using it in practice over the Step-wise SOM. Hence, the purpose of this paper is to propose a GPU parallelization model and implementation for the Batch-SOM using CUDA. The most computationally expensive parts of its training algorithm (such as steps to compute distance between each data vector and neuron, and determining the Best Matching Unit based on minimum distance) are identified and mapped on GPU to be processed in parallel. The proposed implementation shown significant speedups of 11× and 5× compared to the sequential and parallel CPU implementations respectively.
KeywordsSelf-Organizing Map CUDA Clustering Parallel SOM GPGPU
This work is partially supported by Malaysia Fundamental Research Grant Scheme (FRGS) 1/2017/ICT01/UTP/02/2. The experiments reported in this work were performed on the ROMEO computational centre of Champagne-Ardenne, France (http://romeo.univreims.fr). The authors would like to thank J. Loiseau for his useful advices on the GPU implementation.
- 3.Kohonen, T.: MATLAB Implementations and Applications of the Self-Organizing Map, 201 p. Unigrafia Oy, Helsinki (2014)Google Scholar
- 4.Alhoniemi, E., Himberg, J., Parhankangas, J., Vesanto, J.: SOM Toolbox, http://www.cis.hut.fi/projects/somtoolbox. Accessed 15 Jan 2017
- 5.Zhongwen, L., Zhengping, Y., Xincai, W.: Self-organizing maps computing on graphic process unit. In: ESANN’ 2005 - European Symposium on Artifitial Neural Networks, Bruges, Belgium, pp. 27–29 (2005)Google Scholar
- 9.Davidson, G.: A parallel implementation of the self organising map using OpenCL, (2015)Google Scholar
- 10.Kohonen, T.K., Hynninen, J., Kangas, J., Laaksonen, J.: SOM_PAK: The Self-Organizing Map Program Package, 27 p. (1996)Google Scholar
- 11.Wang, H., Mansouri, A., Creput, J.-C.: Massively parallel cellular matrix model for self-organizing map applications. In: 2015 IEEE International Conference on Electronics, Circuits, and Systems (ICECS), pp. 584–587. IEEE (2015)Google Scholar
- 13.NVIDIA: CUDA C Programming Guide (2017)Google Scholar
- 14.Daneshpajouh, H.: NSGA_II-VRP_C101 Dataset, https://drive.google.com/open?id=0ByIxLNivsQhyMVgtN0VFdTRsRms. Accessed 23 Apr 2017