Self supervised learning of EEG (electroencephalogram) raw data to learn the hidden patterns of human brain activities

Abstract

EEG is a non-invasive neuroimaging modality that operates by measuring changes in electrical voltage on the scalp that are induced by cortical activity. In this research, we propose a method for self-supervised learning of EEG raw data to learn the hidden patterns of human brain activities. This work was performed through a pipeline consisting of five phases. Each of the phase’s output will be the input for the next phase. Phase 1 is for pre-processing raw EEG sequences into EEG representations that catch the spacial and temporal properties in the original raw EEG sequences. We have followed a relatively less complex method to pre-process raw EEG sequences. In phase 2, pre-processed raw EEG sequences will be learnt by self-supervised representation learning. For that self-supervised vision transformers with DINO will be used. These vision transformers models are computationally more demanding and require more training data therefore more computational resources and training data will be needed. So that at the presence of more training data and computational processing power, selfsupervised vision transformer architectures will be expected to produce the best results while outperforming supervised learning architectures. Then at the phase 3, sequences of prototypes for each raw EEG data sequence of the dataset will be generated. To evaluate the prototypes that are generated from raw EEG data, phase 4 and 5 have been used as the downstream task for the self-supervised learning task. For phase 4 and 5, we again used a transformer architecture, that is a BERT based model called RoBERTa to learn the synthetic language generated by phase 3 or to learn the context and the language of generated prototype sequences and by performing a multi class prototype sequence classification, prototype generation for each representation at specific time stamp of raw EEG data sequence can be evaluated. We believe that since the models are computationally demanding and require more training data, the latter explained pipeline of five phases should be improved with more training and performing hyperparameter tuning at a high computational resources and data rich environment.