Dual Optimization of Deep CNN for Motor Imagery EEG Tasks Classification

The motor imagery (MI) electroencephalographic (EEG) signals are leveraged as control commands to numerus engineering applications, such as operating a wheelchair through thought alone. EEG signals are characterized by their non-stationary nature and high dimensionality, posing noteworthy challenges for building robust identification models with limited data samples. The renewed deep convolutional neural network (CNN) has been harnessed in various fields due to its ability to autonomously extract and select latent features addressing problems across various types of EEG signals. However, CNN architectures comprise a vast parameter and hyperparameter space. Normally, parameters are optimized via the back-propagation algorithm, while hyperparameters are adjusted through a lengthy trial-and-error process. This work mitigates the aforementioned problems by employing the genetic algorithm (GA) as an artificial intelligence optimization tool for optimizing crucial architectural hyperparameters including the number of convolution filters, size of filters, and dropout probability. Furthermore, the two parameters of cropping augmentation, used for boosting the EEG training samples, are also optimized by the GA. The proposed model in this paper (GACNN) that includes dual optimization, achieved both by the back-propagation algorithm and GA, has attained promising results in the analysis of MI EEG signals. Experiments are performed on the known brain-computer interface competition IV 2a dataset (BCIC IV 2a). GACNN has recorded an increment of about 17% in Cohen’s kappa coefficient and a decrement of about 30% in standard deviation (SD) in comparison to state-of-the-art studies.