Exploring Performance Variability in Deep Neural Networks for Motor Imagery Classification: Towards a Zero-Calibration Brain-Computer Interface

This study focuses on leveraging deep learning and transfer learning techniques for classifying EEG signals recorded during motor imagery tasks. The primary objective is to develop a robust brain-computer interface (BCI) system capable of operating effectively in a zero-calibration scenario. Deep neural networks are investigated to improve classification accuracy. However, performance variability due to uncertainty sources is a critical challenge. To assess this variability, the effects of parameter initialization and pre-processing were analyzed. EEGNet and Sinc-EEGNet served as test models. Results indicate that network's weight initialization significantly impacts performance. Classification accuracy actually improves from 67% ± 3% to 73% ± 3% with adjusted initialization. Conversely, EEG pre-processing does not improve performance. As a result, pre-processing can be omitted to reduce computational effort. These findings support the development of real-time application scenarios.