Abstract
To handle the scarcity and heterogeneity of electroencephalography (EEG) datain Brain-Computer Interface (BCI) tasks, and to harness the vast public data,we propose Neuro-GPT, a foundation model consisting of an EEG encoder and a GPTmodel. The foundation model is pre-trained on a large-scale public EEG dataset,using a self-supervised task which learns how to reconstruct the masked chunkin EEG. We then fine-tune the foundation model on a Motor ImageryClassification task where only 9 subjects are available. Experimentsdemonstrated that applying foundation model can significantly improveclassification performance compared to the model trained from scratch, whichprovides evidence for the advanced generalizability of foundation model and theability to address the challenges of data scarcity and heterogeneity.