A Comparison of Deep Neural Network Algorithms for Recognition of EEG Motor Imagery Signals.

This study aims to compare classical and Deep Neural Networks (DNN) algorithms for the recognition of Motor Imagery (MI) tasks from electroencephalographic (EEG) signals. Four Artificial Neural Networks (ANNs) architectures were implemented and assessed to classify EEG motor imagery signals: (i) Single-Layer Perceptron (SLP), (ii) Fully connected Deep Neural Network (DNN), (iii) Deep Neural Network with Dropout (DNN+ dropout) and (iv) Convolutional Neural Network (CNN). Real EEG signals recorded in a MI-based BCI experiment were used to evaluate the performance of the proposed algorithms in the classification of three classes (relax, left MI and right MI) using power spectral based features extracted from the EEG signals. The results of a systematic performance evaluation revealed not significant classification accuracies with SLP (averaged of 33.9% +/- 0.0%), whereas DNN (59.7% +/- 16.3%), DNN+dropout (58.4% +/- 14.9%) and CNN (62.1% +/- 15.2%) provided significant classification accuracies above chance level. The highest performances were obtained with DNN and CNN. This study indicates potential application of DNNs for the development of BCI systems in daily live activities with real users.