Ensemble deep learning for transfer learning in brain-computer interface

• Grant number: 22/08920-0
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Effective date (Start): September 01, 2022
• Effective date (End): July 31, 2023
• Field of knowledge: Physical Sciences and Mathematics - Computer Science - Computing Methodologies and Techniques
• Principal Investigator: Raphael Yokoingawa de Camargo
• Grantee: Bruna Junqueira de Almeida Ferreira Lopes
• Host Institution: Centro de Matemática, Computação e Cognição (CMCC). Universidade Federal do ABC (UFABC). Ministério da Educação (Brasil). Santo André , SP, Brazil

Abstract

Deep Neural Networks have been successful in Brain-Machine Interfaces (BCIs) applications, where they are used to decode mental images from Electroencephalography (EEG) signals. However, each trained model works for a single individual and requires large amounts of data, which is time-consuming and expensive to collect. Using transfer learning, one can train a neural network on a set of subjects and later adjust its weights to new subjects using small datasets. The main existing approaches are aligning the EEG signals from each subject so that they would have similar characteristics or adjusting parts of the neural network for each subject. In this project, we will evaluate the use of ensemble models for transfer learning. In this case, we will train a neural network for each individual in the training group and create an ensemble with the individual models. We will evaluate different ways of combining the results, including (I) defining weights based on the accuracy of each model on the target subjects and (II) using a meta-model for classification. We will also assess the impact of using a Euclidean alignment of EEG signals. We will use data from motor imagery experiments and compare it with results obtained with other transfer learning techniques.(AU)