Applying CUDA technology for training the Kohonen neural network

The paper presents clustering of the data from in the training samples of the MNIST and Fashion MNIST databases. For clustering, the authors use a Kohonen neural network with a Euclidean metric for estimating distances. The optimal number of clusters (no more than 50) is determined for each handwritten digit (MNIST) and type of clothing (Fashion MNIST). Neural network training is parallelized on a NVidia graphics device using CUDA technology. There are the results for each digit illustrating the comparison of the processor and GPU operating time. For both the digits and clothing types, there is a conclusion about a 17-fold acceleration on an entry-level gaming laptop. Test samples of the same databases are used to verify the cluster construction correctness. For both sequential and parallel learning, it is concluded that the vectors from the test sample belong to the correct cluster with a probability of more than 90 % in the case of handwritten digits. In addition, there are calculated F-measures for each digit and type of clothing to evaluate clusters. It is shown that sequential and parallel clustering give similar results. The best values of the F-measure are obtained for the digits 0 and 1 (F-mean is 0.974), while the worst value is obtained for the digit 9 (F-mean is 0.903). For the Fashion MNIST data, the best value for the F-measure was obtained for trousers (F-average value is 0.96), and the worst value was for a shirt (F-average value is 0.34). Despite the large variations for the F-metric values of the considered two databases, the differences in the clustering results are minimal. Thus, the maximum difference of the F-measure is about 0.01 for the MNIST and about 0.04 for the Fashion MNIST.