Abstract
Serial dependence effects, attractive biases generated by recent visual experience, have been well characterized in behavior, but little is known about their neurophysiological substrate. A recent study by Bell et al. (2020) found that face-gender discrimination bias oscillates at distinct frequencies according to the gender of the previously presented face (13.5 Hz for female, 17 Hz for male). Here, we ask whether this behavioral bias is generated by coherent frequency coding in neural activity of observers. We presented male, female and androgynous faces (17 ms) at different intervals (0.1–1 s) after a phase-resetting auditory stimulus to measure oscillations in gender bias. To evaluate the strength of face representation in memory, we decoded 1-back responses from current EEG activity using support vector machine classifiers. To measure the informativity of different frequency bands, decoding was performed on narrow-band frequency ranges (2 Hz window – 4 Hz to 22 Hz). Observers’ serial dependence and decoding accuracy across frequencies were correlated to identify what signal components encode bias. We confirmed that androgynous stimuli tend to be perceived as the same gender as previously presented faces, and that this discrimination bias oscillates according to previous judgement with respect to a sound cue (14 Hz for female, 18 Hz for male). 1-back responses were successfully classified ~250 ms after stimulus presentation across all frequency bands, but only information at low beta frequencies (14–18 Hz) strongly correlated with observers’ serial dependence (log10(BF10) = 2.9). Activation patterns of classification (EEG activity projected onto decoder coefficients) suggest that the serial bias resides in occipital locations during early visual processing (100–500 ms), before migrating to frontal areas (700–1200 ms). Our results show that behavioral oscillations in face-gender discrimination bias can be traced back to a memory trace in neural activity oscillating in the same frequency range.