Subject-Independent Object Classification Based on Convolutional Neural Network from EEG Signals

• Published in: The ... International Winter Conference on Brain-Computer Interface pp. 1 - 4
• Main Authors: Kalafatovich, Jenifer, Lee, Minji
• Format: Conference Proceeding
• Language: English
• Published: IEEE 22.02.2021
• Subjects: Brain-computer interfaces; convolutional neural network; Electric potential; Electroencephalography; Learning systems; object perception; Pattern classification; Semantics; Statistical analysis; subject-independent learning
• ISSN: 2572-7672
• DOI: 10.1109/BCI51272.2021.9385333

Object classification using brain signals can be widely used in real-life for motor-impaired users. However, brain signal acquisition represents a big challenge for brain-computer interface (BCI) system. Therefore, previous studies have investigated subject independent classification of electroencephalography (EEG) signals using a leave-one subject-out (LOSO) approach as a possible solution. Nevertheless, this results in a huge decrease in classification accuracy. Recent studies analyze EEG signals classification using an adaptation learning approach, which indeed increases classification accuracies compared to LOSO approach; however, it is still unclear the proper amount of data to use and if the results vary significantly depending on the amount of data used. In this work, we aimed to classify EEG signals of observed images into semantic categories. A total of 72 photographs divided into 6 semantic categories were presented to the subjects. Classification was performed using subject independent and adaptation learning methods. Adaptation learning was done using 20%, 50%, and 80% of the target subject's data. Finally, statistical analysis was performed to investigate significant difference between results of used methods. There were significant differences between subject independent and adaptation learning methods; however, there was no statistically significant difference among adaptation learning results. Our results showed the potential of a practical BCI system through the adaptation learning approach.