Attention-based spatio-temporal-spectral feature learning for subject-specific EEG classification

Brain-computer interface (BCI) is a system that recognizes the human intentions from the brain signals for communication with external devices. The electroencephalography (EEG) signals are commonly used for motor imagery based braincomputer interface (MI-BCI) due to non-invasive, cost-effective, and...

Full description

Saved in:
Bibliographic Details
Published inThe ... International Winter Conference on Brain-Computer Interface pp. 1 - 4
Main Authors Ko, Dong-Hee, Shin, Dong-Hee, Kam, Tae-Eui
Format Conference Proceeding
LanguageEnglish
Published IEEE 22.02.2021
Subjects
Bidirectional control
Brain-computer interface
Brain-computer interfaces
Convolutional neural networks
Deep learning
Electroencephalography
Feature extraction
Feature learning
Filter bank
Filter banks
Motor imagery
Self-attention
Subject-specific manner
Online AccessGet full text

Cover

More Information
Summary:Brain-computer interface (BCI) is a system that recognizes the human intentions from the brain signals for communication with external devices. The electroencephalography (EEG) signals are commonly used for motor imagery based braincomputer interface (MI-BCI) due to non-invasive, cost-effective, and portable manner. For the analysis of the EEG signals, there are several machine learning and deep learning methods. However, the majority of those methods have limitations of not considering the distinct frequency bands for subject-specific manner. Therefore, we propose the method that pays attention to the significant frequency bands for each subject and also extracts the spatio-temporal-spectral features simultaneously. We utilize filter bank, sliding window segmentation, and the convolutional neural network (CNN) to extract the spatio-temporal features with consideration of multiple frequency bands. Then, we employ the sub-band attention to determine the significant information of each frequency band. Finally, the attention-based Bi-directional Long-Short Term Memory (Bi-LSTM) is implemented to extract the temporal dynamic features. Our proposed method is evaluated on the BCI Competition IV-2a dataset by using two classes in the subject-specific manner. The experimental results demonstrate that our proposed method is effective to focus on the significant frequency band for each subject.
ISSN:2572-7672
DOI:10.1109/BCI51272.2021.9385293