Deep neural networks for real time Motor-Imagery EEG signal classification

The aim of this research is to develop a high-performance Motor Imagery (MI) classifier capable of using short signal intervals (0.8s) in an effort to move towards real-time performance for Brain-Computer Interfaces (BCIs). First, classification accuracy was investigated with different windowssizes and intervals and compared with baseline levels of performance with common existing methods, Support Vector Machines (SVM) and Linear Discriminant Analysis (LDA), using both spatial and spectral features. It was found that spectral features could produce higher performance using shorter windows compared to spatial features. Next, a state-of-the-art Convolutional Neural Networks (CNN) was developed using the Continuous Wavelet Transformation (CWT), producing a novel Point-wise Convolutional Neural Network (PWCNN) that achieves performance very close to the state-of-theart, namely 80% classification accuracy using the BCI IV 2b dataset operating on 2s intervals; however, random chance performance was found with the BCI IV 2a dataset. Next, to address thelimitations of the PWCNN, a hybrid deep model was developed based on best practice CNNs and Recurrent Neural Networks (RNN). It incorporated novel spatial and temporal attention mechanisms, and is called Convolutional Recurrent Neural Network with Double Attention (CRNN-DA). This model was found to yield 73% classification accuracy and 60% kappa using the BCI IV 2a dataset, which is 3% higher than the winner of the BCI IV 2a competition. A generalisation of the Guided Grad-CAM method suited for EEG signals is also proposed to provide model decision interpretability, which may enable further optimisations to be made. In addition, a novel EEG augmentation technique, to be called shuffled-crossover, is proposed to address the issue of having small datasets for network training. As a consequence of increasing the number of training samples, this approach was found to elicit a further 3% increase in classification accuracy using the CRNN-DA. The suggested model (CRNN-DA) and methods move us closer to realising the aim of practical BCIs capable of responding to multiple input classes in real-time. The proposed double attention mechanism can serve as a feedback loop for data collection, enabling data reflecting user inattention,that may otherwise reduce training efficiency, to be rejected pre-emptively. The proposed augmentation technique can be used to reduce the quantity of training data required. The proposed modified Grad-CAM technique offers an insight into model decisions (viz., model interpretability)that may enable future performance enhancements to be identified more easily.