ИСТИНА

Eye-brain-computer interfaces (EBCIs) are a new class of brain-computer interfaces (BCIs) that are based on on-fly classification of eye gaze behaviors used for sending commands to a computer from spontaneous gaze behaviors [1, 2]. Gaze is a powerful tool for controlling graphical user interfaces, as its fixations can quickly and reliably indicate screen objects which should be clicked, so it can be used instead of a computer mouse by paralyzed people and even, in some cases, by healthy persons. However, spontaneous gaze behaviors used for vision are, in general, not different from those used for control. A BCI might be able to differentiate them, but it would be useful if it could provide acceptable performance on single-trial short segments of brain signal data (a few hundreds of milliseconds). This is a difficult task, given the low signal-to-noise ratio and high variability of the most available brain signal, the electroencephalogram (EEG). Deep neural networks have proven their effectiveness in many fields, such as image classification, speech recognition and natural language processing. In most cases, certain background information from subject area is integrated into the network architecture, such as spatial locality in image processing or sequential structure of data in speech recognition. For the EEG processing, the use of spatial and temporal filters is a reasonable way to incorporate background information. Spatial filters can be considered as pointwise convolution [3] applied in space-time domain. Temporal filters for EEG processing can be applied in different ways, such as depthwise convolutions [4] or 2d convolutions incorporating spatial and temporal information [5]. In the areas where neural networks show most impressive performance, huge labeled datasets are available, so very deep neural networks with plenty of parameters can be used. In the case of BCI, classifier training should be done mainly on small individual datasets, thus the use of huge neural networks would be associated with severe overfitting. Recently, a compact convolutional network, EEGNet, was designed for using in BCIs and demonstrated impressive performance [4, 5]. In the presentation we will demonstrate, for the first time, that its application to the EBCI problem also provides substantial performance improvement, compared to previously used classifiers. We will also discuss more general issues of developing classifiers for EBCIs. This work was supported by the Russian Science Foundation, grant 18-19-00593. References [1] Protzak J. et al. (2013) UAHCI, 662-671. [2] Shishkin S.L. et al. (2016) Front. Neurosci. 10:528. [3] Chollet F. (2016) arXiv:1610.02357. [4] Lawhern V.J. et al. (2018) J Neural Eng. 15(5):056013. [5] Lawhern V.J. et al. (2017) arXiv:1611.08024v2.