Cross-Dataset Motor Imagery Classification with Deep Learning and Riemannian Geometry

• Grant number: 23/00049-1
• Support Opportunities: Scholarships in Brazil - Doctorate
• Start date: May 01, 2023
• End date: April 30, 2026
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computer Systems
• Principal Investigator: Denis Gustavo Fantinato
• Grantee: Lucas Heck dos Santos
• Host Institution: Centro de Matemática, Computação e Cognição (CMCC). Universidade Federal do ABC (UFABC). Ministério da Educação (Brasil). Santo André , SP, Brazil
• Company: Universidade Estadual de Campinas (UNICAMP). Faculdade de Engenharia Elétrica e de Computação (FEEC)
• Associated research grant: 20/09838-0 - BI0S - Brazilian Institute of Data Science, AP.PCPE

Abstract

Brain-Computer Interface (BCI) systems have gained attention recently due to their applications in the medicinal and entertainment fields. Frequently, for these tasks, classification is done from signals recorded from brain activity to enable application control without the usual interactions. However, brain signals such as electroencephalography (EEG) are of high complexity and subject to noise and artifacts, aggravated by the recording systems. The traditional methods, which require manual adjustments and domain knowledge, encounter obstacles due to the large variability of these signals. Lately, to interpret those signals, many authors are using Deep Neural Networks (DNNs) in an end-to-end approach. In this context, EEGNet, built with Convolutional Neural Networks (CNNs), achieved impressive results, both in selective attention and motor imagery paradigms, surpassing the accuracy of traditional methods. In parallel, some works employ Riemannian Geometry (RG) as an alternative, achieving relevant accuracy while training without a gradient optimization method, contrary to DNNs. However, both approaches have limitations, suffering in cross-subject and cross-dataset perspectives. The current research proposal aims to build a hybrid architecture that can take advantage of both strategies. This architecture will be applied in a set of motor imagery datasets, using features from each other, hoping to achieve state-of-the-art performance.