XPySom: Accelerating SOM-based services on multi-cores and GP-GPUs

Self Organizing Maps (SOMs) are unsupervised, shallow, artificial neural networks, built on top of the competitive learning principle, and typically employed for clustering, dimensionality reduction and high-dimensional data visualization. This paper presents XPySom, a high-performance and parallel implementation of SOMs for Python, with a focus on acceleration of the computations through the use of multicore processors and GP-GPUs. After describing the major design and implementation choices, we present results from an extensive experimental evaluation of XPySom on the Extended MNIST openly available dataset of handwritten digits. On this dataset, the training time is shown to be up to two orders of magnitude faster than existing open-source SOM implementations, taking as little as 2 seconds to train the model on the 240000 images in the dataset for 100 epochs. The implementation is released under an open-source GPLv3 license, and made publicly available as the xpysom package for Python.