Cognitive Activity Recognition Based on Self-supervised Learning from EEG Signals

摘要

Identifying the cognitive activities of the human brain is a daunting task due to the fact that those cognitive activities were not observable directly by any known sensing technology. Electroencephalogram (EEG) provided a very useful information that associated with the cognitive activities very closely and was used widely for understanding and recognizing human cognitive activity. In this paper, an experiment was designed to abstractively present the process of "target search" and "action execution" which were among the most common types of cognitive activities during human-computer interaction. EEG signals of the applicants of the experiment were collected and used for subsequent cognitive activity analysis which included a novel temporal-based self-supervised learning approach using BERT to pre-train the data for feature embedding. These encoded points generated by the proposed algorithm could be assigned to the corresponding categories by k-means to capture hidden information about the dominate type of the cognitive activities in any period of 20 ms. The results showed that this approach can distinguish the cognitive activities between the target searching and action executing. And the results suggest that in such a task, subjects' cognitive activities are relatively pure in the initial search moments, but later in the task, multiple activities may be mixed.