ABSTRACTS OF ARTICLES OF THE JOURNAL "INFORMATION TECHNOLOGIES".
No. 3. Vol. 27. 2021

DOI: 10.17587/it.27.125-131

V. V. Sapunov, Junior Researcher, e-mail: wallowind@gmail.com, S. A. Botman, Junior Researcher, e-mail: stepan.botman@gmail.com, G. V. Kamyshov, Engineer, e-mail: gv.kamyshov@mail.ru, N. N. Shusharina, Ph.D., Head of Research Organization Department, e-mail: nnshusharina@gmail.com, Immanuel Kant Baltic Federal University, Kaliningrad, 236016, Russian Federation

Application of Convolution with Periodic Boundary Condition for Processing Data from Cylindrical Electrode Arrays

In this paper, modification of convolutional neural networks for purposes of processing electromyographic data obtained from cylindrical arrays of electrodes was proposed. Taking into account the spatial symmetry of the array, convolution operation was redefined using periodic boundary conditions, which allowed to construct a neural network that is invariant to rotations of electrodes array around its axis. Applicability of the proposed approach was evaluated by constructing a neural network containing a new type of convolutional layer and training it on the open UC2018 DualMyo dataset in order to classify gestures basing on data from a single myobracelet. The network based on the new type of convolution performed better compared to common convolutions when trained on data without augmentation, which indicates that such a network is invari­able to cyclic shifts in the input data. Neural networks with modified convolutional layers and common convolutional layers achieved f-1 scores of 0.96 and 0.65 respectively with no augmentation for input data and f-1 scores of 0.98 and 0.96 in case when train-time augmentation was applied. Test data was augmented in both cases. Potentially, proposed convolution can be applied in processing any data with the same connectivity in such a way that allows to adapt time-tested architectural solutions for networks by replacing common convolutions with modified ones.
Keywords: artificial neural network, convolutional neural network, convolution, equivariance, invariance, cyclic shift, electromyography, electrodes array

P. 125–131

To the contents