Solving the Inverse Problem for EEG Signals When Learning a New Motor Task Using GRU Neural Network

Electroencephalogram (EEG) is a noninvasive technique for recording brain neural activities. It has a poor spatial resolution compared to its temporal resolution. However, the inverse problem has to be solved to find neural sources of brain activity. In recent years artificial neural networks have been increasingly used for solving EEG inverse problem. In these methods, source reconstruction is mostly done sample by sample, while the neural sources are highly interconnected. To consider the temporal dependencies, in this research, a neural network structure based on GRU is presented, which has a low computational cost and is resistant to noise. In this novel structure, GRU networks can extract spatial and temporal information from EEG signals. Also, we employ an encoder-decoder structure which learns a latent-space representation to denoise data. Using simulated data, it has been shown that the presented method performs better than the classical methods on several defined criteria, such as AUC, MLE, and nMSE. Then the trained model was used to solve the inverse problem for real EEG data collected during a new motor task while drawing some shapes with the dominant leg.