Fig. 4: Localizer task trial structure, behavioral results, univariate fMRI analyses and training of machine learning-based face-processing model.

a The 1-back localizer task had comparable presentation times as chosen in the feedback-based association learning task for the stimulus presentation and the inter-trial-interval. In two of 16 trials per run, direct stimulus repetitions occurred. On these repetitions, participants (n = 30, 15 male, 15 female) were instructed to quickly press the confirmation key. The task consisted of five runs, each containing two new faces and two new houses. b Most participants performed the task without mistakes (misses or false alarms). Two participants were excluded from further fMRI analyses of the localizer task and later multivariate pattern analyses, because task comprehension and attention to the task could not be assured. c Conventional general linear model analyses showed that hemodynamic responses were larger for faces than houses in the right fusiform gyrus, and larger for houses than faces in the parahippocampal gyrus (PHG). The fusiform gyrus cluster largely overlapped with a cytoarchitectonic probability mask for fusiform gyrus 4 (cFG-4) and was identified as fusiform face area (FFA). The color bar indicates z-scores and the results are thresholded using a false discovery rate (q < 0.05). d A probability-scaled linear support vector machine was trained to distinguish faces and house based on the 14 strongest voxels in the cFG-4 mask according to analysis of variance feature selection, to quantify FFA-based face-processing evidence. Average classification accuracies during leave-one-run-out cross-validation showed no statistical difference in the prediction for faces and houses. e The assessed distance from the multivariate hyperplane indicates evidence for face-processing as shown in the schematic overview.