Hybrid Convolutional, Recurrent and Attention-Based Architectures of Deep Neural Networks for Classification of Human-Computer Interaction by Electroencephalography

Nikita Gordienko, Oleksandr Rokovyi, Yuri G. Gordienko, Sergii G. Stirenko

Abstract: Nowadays, applications for monitoring physical activity and health are becoming popular, especially for human-computer interaction (HCI) by users with some physical disabilities. Electroencephalography (EEG) data analysis of some HCI-related activities can be useful to support everyday life of such people. Recently several approaches based on artificial intelligence methods, like neural networks (NN), for example, fully connected NN (FCN), convolutional NN (CNN), recurrent NN (RNN), were successfully used for EEG data analysis. Some new attention-based NN (wA) architectures are very promising in various applications. This work is dedicated to the investigation of various hybrid combinations, like FCN-CNN, CNN-RNN, CNN-wA, RNN-wA, CNN-RNN-wA, etc. with regard to EEG data analysis. These hybrid models were trained on the grasp-and-lift (GAL) dataset where users use their arm to manipulate a smartphone.