트랜스포머 기반 2-Phase 학습법을 통한 EEG 데이터 분류 모델 성능 개선

다국어 초록 (Multilingual Abstract)

Electroencephalogram (EEG) brain signals provide a wealth of information, and Brain-Computer Interface (BCI) technology leveraging these signals offers hope for individuals with disabilities. Recent research has explored various fields with EEG signal like Motor Imagery, Emotion recognition etc. Among them, there are several studies on object imagery to synthesize object class from brain signals. However, previous work has demonstrated poor performance in EEG classification. This is because EEG signals have different characteristics from person to person, making them difficult to generalize, and it led to itself in differences between train and test performance. In our study, we address this limitation by employing a Transformer-based approach and a two-phase training strategy, leading to improved classifier performance. Consequently, our enhanced classifier show better performance and less gap between train and test performance. Also when we visualize the features for each class, we can see that they are organized into a better refined distribution by label. And we can expect that this can contribute to various applications including assistive technologies for individuals with disabilities.