Spatiotemporal Convolutions on EEG signal: A Representation Learning Perspective on Efficient and Explainable EEG Classification with Convolutional

Deep learning using convolutional layers along the temporal and spatial dimensions is a prevalent and successful way to perform classification on EEG signals, with applications in many different fields. Most of these models use two independent one-dimensional convolutional layers concatenated. In this paper, we investigate an alternative representation that operates a bi-dimensional spatiotemporal convolution. Through a series of empirical tests, we can observe that this new representation leads to significantly reduced training time in a shallow CNN and a more advanced conformer model. Furthermore, an analysis of the activation patterns to establish the difference between the 1D and 2D convolutional layers shows that while the learned spectral features are consistent across different model types, the representations of these features are significantly different. Overall, we suggest an improved model using a 2D convolutional layer for improved training and inference speed. We highlight the importance of training speed for faster development hopefully leading to more robust results and fast inference speed for the viability of deep learning models in online BCI applications.