Decoding the sex of faces using the power, phase, and the Fourier spectrum of the time-frequency representation
The EEG activity related to face processing has been studied extensively with machine learning techniques and most of these studies apply preprocessed data without further data transformation. Here we analyzed the data of two experiments and explored the potential in decoding the time-frequency representation in face processing, thus extracting not only temporal but also the frequency distribution. In the two experiments, participants were presented with faces and through changing the presentation number per face, we manipulated their familiarity. Then we determined the time window of facial sex related cortical activities with frequently employed decoding techniques on the preprocessed data. Subsequently, we performed Fourier transformation on the data and decoded time-frequency spectrum of the amplitude, the phase and also the complex Fourier spectrum. This analysis revealed a 500 ms long time window at the beginning of the stimulus presentation in the 2 to 17 Hz frequency range, which showed above-chance decoding accuracies in the case of more familiar faces. Less familiar faces showed similar, albeit more restricted time-frequency windows. By comparing the two experiments we also observed 350 ms long window in the low frequencies of 4-10 Hz, where familiar faces exhibited greater decoding accuracies. This method expanded on the generally observed time-window and complemented it with a frequency distribution related to facial sex processing. The current study demonstrates that machine learning applications can be applied to higher dimensional data, like the time-frequency representation in cognitive studies.