Identifiable EEG Embeddings by Contrastive Learning from Differential Entropy Features

Encoding EEG data into low-dimension latent embeddings greatly facilitates data analysis and interpretation in neuroscience studies, clinical diagnosis, and human-computer interaction. But generating informative and identifiable latent embeddings that are representative of the origin EEG is not an easy mission. Contrastive learning has the potential to utilize large amounts of unlabelled EEG data and extract informative and identifiable latent embeddings for a wide range of downstream tasks. We explore the feasibility of applying the contrastive learning method to train the EEG latent encoder from the feature of differential entropy of short-time window frequency domain signals. The encoder minimizes the noise-contrastive estimation loss by comparing the embeddings with positive and negative embedding samples, where the distinction of samples is guided by time nearness information or task-specific labels. We test encoders with different output dimensions and the outcome latent embeddings can be identifiable via visualization of a few dimensions. The decoding result also shows that the embeddings preserve information about the original EEG features and can be potentially used for a wide range of downstream tasks. The source code is available at: https://www.github.com/liangfengsid/deContrastiveLearning.