Unconditional EEG Synthesis Based on Diffusion Models for Sound Generation

Classifiers used in brain-computer interfaces based on the electroencephalography (EEG) typically demonstrate rel-atively low performance, which is a serios obstacle for making them a practical technology. One of the most important limitations that prevents improving EEG classification is the scarcity of the EEG data. Thus, generation of synthetic data could help to enhance classification. Recently, diffusion models were applied time for time series generation and first steps were made in generating synthetic EEG data using them. Here, we introduce MultiChan Wavegrad, a novel diffusion model designed specifically for multichannel EEG data generation. We describe its architecture and preliminary results of its testing using the BCI competition IV 2a dataset with the EEG recorded during motor imagery. The data generated by MultiChanWaveGrad possessed some resemblance to the real EEG data, although did not reproduce the EEG characteristics well enough. Finally, we discuss possible future directions for improving its performance and possibly making it a useful tool for data augmentation, especially for improving training of BCI classifiers.