aweSOM - Accelerated Self-organizing Map and Statistically Combined Ensemble
The aweSOM package combines a JIT-accelerated and parallelized implementation of self-organizing map (SOM), integrating parts of POPSOM, and a GPU-accelerated implementation of statistically combined ensemble (SCE) using ensemble learning.
aweSOM was developed for machine-learning clustering and classification tasks, with a focus on identifying intermittent structures in 3D plasma simulations (link to arXiv pre-print). Past implementations of SOM and SCE were single-threaded and not optimized for large datasets, so aweSOM was developed to address this issue. Using a combination of JIT-accelerated and parallelized SOM and GPU-accelerated SCE, aweSOM can handle datasets with up to \(\sim 10^8\) points running on a single GPU/CPU node.
Additionally, aweSOM has been designed to be general-purpose, so it can be used for a variety of clustering and classification tasks beyond its original purpose. See the notebook for an example application on the classic Iris dataset.
Authors:
Trung Ha - University of Massachusetts-Amherst, Joonas Nättilä - University of Helsinki, and Jordy Davelaar - Princeton University.
Current version: 1.1.0