Motor imagery is currently one of the main applications of Brain-Computer Interface (BCI) which aims at providing the disabled with means to execute motor commands. One of the major stages of motor imagery systems is reducing the dimensions of the input data and enhancing the features prior to applying a classification stage to recognize the intended movement. In this paper, we utilize autoencoders as a powerful tool to enhance the input features of the band power filtered electroencephalography (EEG) data. We compare the performance of the autoencoder-based approach to using Principal Component Analysis (PCA). Our results demonstrate that using autoencoders with non-linear activation function achieves better performance compared to using PCA. We demonstrate the effects of varying the number of hidden nodes of the autoencoder as well as the activation function on the performance. We finally examine the characteristics of the trained autoencoders to identify the features that are most relevant for the motor imagery classification task.
