Mobile Sensor Data Classification Using GM-SOM

Abstract

The paper uses a previously-introduced modification of standard Kohonen network (SOM), called GM-SOM. This approach uses partitioning the problem in case of insufficient resources (memory, disc space, etc.) and parallel processing of input data set to process all input vectors at once, with the use of modern multi-core GPUs to achieve massive parallelism. The algorithm pre-selects potential centroids of data clusters in the first step and uses them as weight vectors in the final calculation. In this paper, the algorithm has been demonstrated on a new UCI HAR dataset, representing activities recorded by smartphone sensors, which are prone to random noise due to the sensor behavior. Moreover the separation of classes is not linear, which introduces additional complexity and makes it hard to process the data by linear algebra methods. © Springer International Publishing Switzerland 2013.