Figure 3

Panels left: The three filters generated input data with maximum activation for layer 2 (top), layer 3 (middle) and layer 6 (bottom), representing the features with most discriminative power between sexes. The Fourier Spectral analysis in the time domain (panels right, averaged for all 24 rows, similar to a traditional Fourier analysis approach of EEG data) displays that filters of layer 2 yield white noise without prominent features, while filters for layer 3 show a distinct peak in the beta range (12â25âHz), similar to layer 6. In this example, the activation of layer 6 is mainly seen within rows (i.e. EEG-channels) 4â20, restricting the extracted feature to certain EEG-channels, providing additional spatial information.