FIGURE 1.

Sex detection with the Mini‐SpaTeN convolutional neural network and relevance attribution method. First, the segment $\left(X\right)$ is cross‐correlated $\left(\star \right)$ with 16 learned spatio‐temporal kernels ($K_i$) similar in dimension as short windows of EEG (actual kernels depicted in Figure 2). Since the kernels have the same number of channels as the data, they only slide along the time axis and not across channels. The 16 correlation curves are rectified (ReLU activation) and divided into 40 overlapping windows. Next, the window‐wise maxima ($M_{\mathit{ij}}$) are averaged. In the final layer, the sex $\left(Y\right)$ is predicted from these 16 averages with logistic regression. Post hoc, the network parameters are used to attribute the relevance $R$ that each EEG channel and time point from a recording had to the prediction (path indicated in purple). The signs of the final classifier layer weights ($w_i$) imply the sex corresponding to each kernel from the first layer (−/female/red and +/male/blue).