Foundations of Conditional Diffusion Models for Event-Related Potentials
Keywords
Loading...
Authors
Issue Date
2024-08-21
Language
en
Document type
Journal Title
Journal ISSN
Volume Title
Publisher
Title
ISSN
Volume
Issue
Startpage
Endpage
DOI
Abstract
Generative models, specifically diffusion models, can alleviate data scarcity in
the brain-computer interface field. Moreover, they could potentially offer a
promising solution to calibration sessions. While diffusion models have previ ously been successfully applied to electroencephalogram (EEG) data, existing
models lack flexibility regarding sampling and often require alternative represen tations of the EEG data. To overcome these limitations, we introduce a novel
approach to conditional diffusion models that utilizes classifier-free guidance
to directly generate event-related potential (ERP) EEG data that is specific
to a combination of labels in the dataset. Moreover, we evaluate the model’s
ability to generate samples for label combinations excluded during training,
demonstrating its potential for transfer learning. In addition to commonly
used metrics, domain-specific metrics are introduced to evaluate the specificity
and quality of the generated samples. The results indicate that the proposed
model can generate ERP EEG data that resembles real data for each combi nation of labels in each dataset. Furthermore, the model is capable of within session between-class and between-class transfer learning, while between-session
transfer learning remains elusive. The code used for this work is released at:
https://gitlab.socsci.ru.nl/neurotech/code/thesis_guido_klein
Description
Citation
Faculty
Faculteit der Sociale Wetenschappen