FIGURE 2.

Learned male and female weights extracted from the trained CNN. (a) From the 16 weights $w$ of the final classification layer, half individually converged to a positive and half to a negative sign, which determines the sex, related to the corresponding kernels below. Female weights are marked red, male weights are blue. The spatio‐temporal kernels $K$ are similar in dimension to short segments of EEG, but may not be interpreted as EEG (Haufe et al., 2014). As the CNN searches for strong cross‐correlations between the kernels and the data, the kernel values can have three meanings: large kernel values with the same sign as the data are positively added to the correlation sum, small kernel values ignore the underlying data, and inverse values reduce the correlation sum and therewith the relevance of the whole segment. Thus, the kernel waveforms can also show anti‐patterns, instead of the searched distinctive activity. Part of a kernel can be concurrently correlated with sex differences and just be used for localization in the data, but not be sex‐different itself. (b) Weights $w$ and single‐timepoint kernels $K$, trained on temporally shuffled data. Figure S4 shows a variant of (a) in which the spatio‐temporal kernels were trained on raw data instead of ICA‐filtered data. They show waveforms that more strongly resemble the cardiac QRS complex.