Transposed Convolution as Alternative Preprocessor for Brain-Computer Interface Using Electroencephalogram

Abstract

The implementation of a brain–computer interface (BCI) using electroencephalography typically entails two phases: feature extraction and classification utilizing a classifier. Consequently, there are numerous disordered combinations of feature extraction and classification techniques that apply to each classification target and dataset. In this study, we employed a neural network as a classifier to address the versatility of the system in converting inputs of various forms into outputs of various forms. As a preprocessing step, we utilized a transposed convolution to augment the width of the convolution and the number of output features, which were then classified using a convolutional neural network (CNN). Our implementation of a simple CNN incorporating a transposed convolution in the initial layer allowed us to classify the BCI Competition IV Dataset 2a Motor Imagery Task data. Our findings indicate that our proposed method, which incorporates a two-dimensional CNN with a transposed convolution, outperforms the accuracy achieved without the transposed convolution. Additionally, the accuracy obtained was comparable to conventional optimal preprocessing methods, demonstrating the effectiveness of the transposed convolution as a potential alternative for BCI preprocessing.