Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data Augmentation

Poļaka, Agnese Viktorija (2025) Riemannian Geometry-Preserving Variational Autoencoder for MI-BCI Data Augmentation. Bachelor's Thesis, Artificial Intelligence.

Preview	Text AIBachelorsThesisFINAL.pdf Download (2MB) | Preview
	Text Toestemming.pdf Restricted to Registered users only Download (183kB)

Abstract

This paper addresses the challenge of generating synthetic electroencephalogram (EEG) covariance matrices for Motor Imagery Brain-Computer Interface data augmentation. Objective. To develop a generative model capable of producing high-fidelity synthetic covariance matrices for BCI data augmentation, accounting for the Symmetric Positive-Definite constraint. Approach. A novel Riemannian geometry-preserving Variational Autoencoder architecture is proposed. This model integrates geometric mappings and employs a composite loss function that combines Riemannian distance for manifold fidelity with objectives promoting tangent space reconstruction and generative diversity. Results. The model successfully generates valid and representative EEG covariance matrices. The utility of the synthetic data was evaluated in a cross-subject, Leave-One-Subject-Out Cross-Validation classification setting and found to be highly classifier-dependent. While the augmentation significantly hindered the performance of a Support Vector Classifier, it maintained performance using Minimum Distance to Mean classifier and even provided a statistically significant improvement for the geometry-aware K-Nearest Neighbors classifier, increasing its balanced accuracy by up to 3.49%. Contribution. This work validates a new architecture for generating Motor Imagery EEG covariance matrices and concludes that its effectiveness is directly linked to the algorithm of the classifier it is paired with.

Item Type:	Thesis (Bachelor's Thesis)
Supervisor name:	Jong, I.P. de and Sburlea, A.I.
Degree programme:	Artificial Intelligence
Thesis type:	Bachelor's Thesis
Language:	English
Date Deposited:	05 Aug 2025 12:21
Last Modified:	05 Aug 2025 12:21
URI:	https://fse.studenttheses.ub.rug.nl/id/eprint/36690

Actions (login required)

View Item