Meta-Learning for Subject Adaptation in Low-Data Environments for EEG-Based Motor Imagery Brain-Computer Interfaces

Arnav Pati, Deepak Mewada, Debasis Samanta

Keywords: neuroscience, brain computer interface, EEG, meta learning, transfer learning, application of deep learning

Abstract: Motor imagery classification from Electroencephalogram (EEG) signals involves decoding information during the imagination of specific movements. However, learning representations for EEG-based motor imagery classification is challenging due to inter-subject variability and differences in mental imagery, resulting in poor generalization of deep learning models to new subjects. While pre-trained deep learning models achieve high accuracy on subjects with similar domains, they fail on subjects with dissimilar domains. Optimization-based meta-learning algorithms can address this limitation by learning a good initialization for the model, enabling quick adaptation to new subjects with limited fine-tuning examples. We demonstrate that our Meta Learning approach consistently outperforms Transfer Learning on the BCI Competition IV 2a dataset. Although accuracy varies depending on domain similarity, meta-learning demonstrates efficient adaption to unseen subjects with limited data. By improving generalization across subjects with different domains under low-data environments, we can enhance the reliability and practicality of brain-computer interfaces for real-world applications.