Investigating the EEG Embedding by Visualization

Visualizing EEG data helps clinical doctors and neuroscientists discover potential patterns and abnormalities before further mathematical analysis. Encoding complex EEG data into low-dimension embeddings and visualizing the points in 3-dimension axes with colors can help users quickly recognize some EEG properties. We apply contrastive learning in both self-supervised and supervised manners to extract the time-domain EEG features within different time window sizes. The color points tend to cluster into clouds based on their related classes and graph readers can roughly distinguish people’s emotions and identities directly by inspecting the graphs. With self-supervised encoders where the generated embeddings are supposed to be used for general tasks, the visualization method can also uncover the value of the original input features extracted from raw EEG data. The source code is available at: https://www.github.com/liangfengsid/visContrastive .