Vsom: Efficient, stochastic self-organizing map training
Document Type
Conference Proceeding
Date of Original Version
1-1-2018
Abstract
Here we introduce VSOM, an efficient implementation of stochastic training for self-organizing maps. We derive VSOM from the standard stochastic training algorithm as published by Kohonen by replacing all iterative constructs in the algorithm with vector and matrix operations. Our novel implementation based on these vector and matrix operations provides substantial performance increases over Kohonen’s iterative algorithm as well as batchSOM, currently the fastest implementation of Self-organizing maps (SOM) training without resorting to multi-processing. The quality of the maps produced by VSOM matches the quality of the maps produced by the original iterative algorithm and outperforms the quality of the maps produced by batchSOM. In its current incarnation VSOM is single threaded and therefore well suited as a replacement for iterative stochastic training of self-organizing maps in R since R does not support multi-threading well.
Publication Title, e.g., Journal
Advances in Intelligent Systems and Computing
Volume
869
Citation/Publisher Attribution
Hamel, Lutz. "Vsom: Efficient, stochastic self-organizing map training." Advances in Intelligent Systems and Computing 869, (2018): 805-821. doi: 10.1007/978-3-030-01057-7_60.