Gate Control Mechanisms of Autoencoders for EEG Signal Reconstruction

Electroencephalography (EEG) is a non-invasive and portable way to capture neurophysiological activity, which provides the basis for brain–computer interface systems and more innovative applications, from entertainment to security. However, the acquisition of EEG signals often suffers from noise contamination and even signal interruption problems due to poor contact of the electrodes, body movement, or heavy noise. Such heavily contaminated and lost signal segments are usually removed manually, which can hinder practical system deployment and application performance, especially in scenarios where continuous signals are required. In our previous work, we proposed the weighted gate layer autoencoder (WGLAE) and demonstrated its effectiveness in learning dependencies in EEG time series and encoding relationships among EEG channels. The WGLAE adopts a gate layer to encourage the AE to approximate multiple relationships simultaneously by controlling the data flow of each input variable. However, it only applies a sequential control scheme without taking into account the physical meaning of EEG channel locations. In this study, we investigate the gating mechanism for WGLAE and validate the importance of having a proper gating scheme for learning relationships between EEG channels. To this end, several gate control mechanisms are designed that embed EEG channel locations and their corresponding underlying physical meanings. The influences introduced by the proposed gate control mechanisms are examined on an open dataset with different scales and associated with various stimuli. The experimental results suggest that the gating mechanisms have varying influences on reconstructing EEG signals.