Abstract
Brain-Computer Interface (BCI) systems have a great impact on improving people’s lives. One of the popular BCI implementations is the Event-Related Potential (ERP)-based spelling system which decodes electroencephalogram (EEG) signals to identify a target character. The effectiveness of BCI systems highly depends on the single trial decoding accuracy; however, the EEG signals are contaminated with diverse artifacts which leads to a poor signal-to-noise ratio. Therefore, various filtering algorithms (e.g., FFT, CSP, Laplacian, PCA) have been applied to find the optimal subset of feature spaces in the temporal and spatial domains. These preprocessing steps could efficiently discard the artifacts and have shown superior performance with typical linear classifiers. However, there is a risk that the informative subspace can be also eliminated by the unsupervised learning process, and this algorithm is not proper to be employed in the end-to-end deep-learning architectures where all modules are differentiable. This study aims to propose a generalized deep neural network model by denoising the ERP signals and initializing the Convolutional Neural Network (CNN) model parameters based on the autoencoder. Proposed CNN models indicate - 98.2% spelling performance and - 91.5% single trial accuracy which outperformed the state-of-the-art CNN models.