New avenues for understanding what deep networks learn from EEG

An important but unresolved question in deep learning for EEG decoding is which features neural networks learn to solve the task. Prior interpretability studies have mainly explained individual predictions, analyzed the use of established EEG features, or examined subnetworks of larger models. In contrast, we apply interpretability methods to uncover features learned by the complete network. Specifically, we introduce two complementary architectures with dedicated visualization techniques to obtain an approximate understanding of the full network trained on binary classification into nonpathological and pathological EEG. First, we use invertible networks-networks that are designed to be invertible-to generate prototypical input signals for each class. Second, we design a very compact network that is fully visualizable, while still retaining reasonable decoding performance. Through these visualizations, we find both expected features like higher-amplitude oscillations in the delta and theta frequency bands in the temporal region for the pathological class as well as surprising differences in the very low sub-delta frequencies below 0.5 Hz. Closer investigation reveals higher spectral amplitudes for the healthy class at the frontal sensors in these sub-delta frequencies, an unexpected feature that the proposed visualizations helped identify. Overall, the study shows the potential of visualizations to understand the network prediction function without relying on specific predefined features.