Our goal here was to investigate whether face templates support recognition of Mooney faces, and whether the templates, if present, are observer-specific. We used classification images to visualize observers’ face templates. To generate the classification images, we averaged all of the observer's selected noise patterns across trials, and we superimposed the resulting noise on the base face (
Figure 2C). The classification image reveals the regions or structures that observers rely on to classify an image as face-like. We generated one classification image per base face, per subject. The outcome was eight unique classification images for each observer.
To evaluate whether observers’ classification images were nonrandom, we tested the day-to-day consistency of each observers’ classification images for each of the eight base Mooney faces. The within-subject consistency was defined as the test-retest pixel-wise correlation between an observer's classification images in the first and second session. We then averaged all within-observer Fisher z transformed correlations for upright and inverted base faces, separately. Second, we quantified the between-observer agreement in classification images of Mooney faces. To this end, we calculated the between-observer correlations across different observers’ classification images for each base face. Then we averaged all between-observer Fisher z transformed correlations for upright and inverted base faces, separately.
To test the significance of the within- and between-subject correlations, we calculated a null distribution of permuted correlations. In each iteration of the permutation, for each base face, we shuffled the responses that were given by each participant. We then calculated a classification image from these shuffled responses. This was effectively the classification images an observer would have if the participant responded randomly. Using this data, we generated a set of 1000 permuted classification images, per base face, per participant. Note that our classification images have signal included, which could artificially inflate the correlations in both within- and between-subject analyses. To control for this, we calculated null distributions of permuted correlations, which maintain all of the same signal information, but represent shuffled responses, so any inflation of the correlations will occur in the null distribution as well. In other words, we never compare correlations to zero, always to permuted shuffled null distributions. To generate the distribution of permuted null within-subject correlations, for each observer and each base face, we correlated the empirical classification image of the first session with each of the permuted null classification images of the second session. In the same fashion, to generate the distribution of permuted null between-subject correlations, for each pair observers and each base face, we correlated an observer's one-session empirical classification image with each of the permuted null classification images of another observer's session.
Last, we quantified the individual differences of classification images of Mooney faces by comparing within- and between-observer agreement. The significance of this comparison was tested using a nonparametric permutation method: we computed the empirical difference for within- and between-correlations and compared it to the null difference for the between and within correlations for randomized responses. Across all analyses, to calculate the average within- and between-subject correlations across base faces or across observers, correlations were first transformed from Pearson R correlations into Fisher z correlations before averaging them.