Seventeen neurologically healthy male subjects between the ages of 25 and 41 (M = 33, SD = 4.8), with normal or corrected vision participated in the experiment. Only males were used because of limitations on the pipeline to generate look-alike avatars. Participants were recruited via email from the laboratory mailing list. Since we wanted to study the effects of self-specific processing and familiarity, participants were matched with workmates and friends seen every day by the participants for the stimuli generation. Participants had never been exposed to the specific stimuli used in this experiment at the moment of recruiting and had not participated before on face observation studies. Subjects gave informed consent according to the Declaration of Helsinki, and the experiment was approved by the ethics committee of the Universitat de Barcelona.

We took three photographs (front, left and right profile) of each participant in order to create their look-alike avatars. In all cases participants were requested to maintain a neutral facial expression to avoid emotion processing during the observation, since it has been shown that emotions on faces generate different brain traces and activate mirror neurons (Likowski et al., 2012). Avatars were generated in a matter of minutes using the fast creation of look-alike avatars pipeline described in Blom et al. (2014). The resultant avatars were tweaked manually to make minor smoothing and lighting improvements. The avatars were displayed without hair, focusing the experiment on the participant’s pure facial characteristics; none of the participants was actually bald. We used the original frontal picture and a frontal capture of the resulting avatar for the observation task (Figure 1A).

The real and the computer generated images were processed to equalize pupil-pupil distance across all, and to ensure that the vertical midline of the image bisected the face; this involved a minor scaling of the entire image, and left the normal proportions of the face invariant. The images were saved at 520 × 520 pixels, in color and with averaged luminance to avoid effects of uncontrolled interstimulus perceptual variance that have been found to influence the N170 component (Thierry et al., 2007). Controlling this particular variance seems to correct the N170 face sensitivity (Bentin et al., 2007).

The experimental design included two factors: Virtuality and Familiarity. The Virtuality factor consisted of either a photographic image of a real person (Real) or an image of the look-alike avatar (Avatar) of that person. The Familiarity factor consisted of the three levels: self (Self), familiar (Familiar) or unfamiliar other (Unfamiliar). The person used for the Unfamiliar condition had no previous relation to any of the participants. Since we wanted to study the effects of self-specific processing rather than habituation to the face—e.g., participants see themselves in mirrors very frequently—the Familiar faces were extracted from friends seen every day by the participants.

In total, subjects were exposed to the six images, 200 times each, in a classical ERP setup resulting in a total of 1200 faces to be observed by each participant during the experiment that lasted approximately 30 min. Habituation to the stimuli was explored to test possible adaptation effects. The Familiarity factor was based on the existing literature on ERPs and self-face processing (Sui et al., 2006; Tanaka et al., 2006; Keyes et al., 2010).

The visual stimuli were implemented and displayed using the XVR programming system (Tecchia et al., 2010; Spanlang et al., 2014) on an Intel Core i7 at screen resolution 1920 × 1080 pixels. The six faces (self-real, self-avatar, friend-real, friend-avatar, unfamiliar-real, unfamiliar-avatar) were pseudo-randomly ordered in blocks of 10. Each face was displayed for 300 ms, followed by a variable time of 740–1340 ms in which a fixation cross appeared (Figure 1B). After each block of 10 pictures there was a resting period of 2 s for blinking. Participants were instructed to maintain the focus on the fixation point and to minimize blinks and eye movements.

Before mounting the EEG cap and starting the experiment we showed the faces one by one to the participants and administered a short questionnaire in which they had to rate the realism of both the real and avatar pictures from 1 (not alike at all) to 5 (looks completely like the real person). Specifically they were asked to answer whether: “The picture looks like the real person”. The realism of the faces when compared to the real person were not available for the unfamiliar case as participants did not know the real person. Instead they were asked to compare the similarity of the avatar face to the photographic picture face (from 1 to 5): “The avatar face looks like the real face”. This similarity rating was performed for all pairs of real-avatar pictures. In order to detect changes in the scoring that were due to the experimental exposure or adaptation effects, participants were asked to rate again all the faces and avatars after the experiment.

Continuous EEG was acquired from 64 active electrodes located at standard 10–20 positions with a g.HIamp multichannel amplifier manufactured by g.Tec Medical Engineering. Active ring electrodes (g.LADYbird) were used in a standardized cap (g.GAMMAcap), both from g.Tec. The activity was referenced to the earlobe and the ground electrode was located in the frontal area of the head. Signals where digitalized at 256 Hz frequency rate, a notch Butterworth filter 4th order from 48 to 52 Hz was used to eliminate the AC. Ocular movements were detected from FP1, FP2, AF7 and AF8.

The EEG was segmented offline into 1200 ms epochs starting 200 ms before the stimulus onset. Trials where the faces were consecutively repeated were rejected off-line. There was a repetition rate of 2.99%. Trials in which ocular movements were found (EOG greater than 50 μV) or the absolute amplitude of the signal at any electrode was greater than 150 μV were rejected. The average acceptance rate was of 75 ± 15% trials per participant.

ERPs time-locked to the onset of the stimuli for each condition and participant were averaged for epochs of −200 to 900 ms with the baseline set from −150 to 0 ms. Mean amplitudes were calculated where there were discernible peaks in the average ERP waveform for each of the 17 participants; these included the epochs N170 (160–220 ms) and P200 (250–300 ms) of the right and left occipito-parietal cortex, which are consistent with those proposed in previous self-recognition experiments (Sui et al., 2006; Tanaka et al., 2006; Keyes et al., 2010).

The original ERPs for each subject were also transformed into reference-free Current Source Density (CSD) estimates (μV/cm² units) using a spherical spline surface Laplacian Matlab-based CSD toolbox (Kayser and Tenke, 2006) to better show the topographical maps.

Participants rated of the realism of the faces by responding to the statement: “The face looks like the real person”; and the similarity between the avatar face and the photograph by responding to: “The avatar face looks like the face in the photograph” (Figure 2). After the experiment, participants were asked to rate again all the faces. There were no significant differences between any of the real or virtual faces in the pre-post scoring for the realism question (Wilcoxon signed-rank tests, all p > 0.45, n = 17), nor for the similarity pre-post (Wilcoxon signed-rank tests, all p > 0.55, n = 17). The real image scores served as a consistency control: none of the real pictures scored less than 4 out of the maximum of 5. Furthermore, there were no significant differences amongst the scores for all the avatar faces, which indicates that all of them were perceived similarly realistically (Friedman test, p > 0.37, n = 17).

There were significant differences when comparing the realism scores between the avatar and the real faces (Wilcoxon signed rank tests, all p < 0.001). This indicates that even though all the participants did recognize themselves and their confederates in the avatar faces, the avatars were not perceived as overly realistic as the real faces (Figure 2A). To the similarity question all avatars (self, familiar and unfamiliar) were rated as equally similar to their real counterparts (Figure 2B).

In order to explore how the faces were processed in the brain we proceeded to study well-defined visual processing-related ERP components in occipito-parietal cortex, such as the N170 and the P200 components (Bentin and Deouell, 2000; Caharel et al., 2002; Sui et al., 2006; Tanaka et al., 2006; Keyes et al., 2010). Visual inspection of Figure 3 shows a clear negative voltage for all the faces during the N170 (both for avatar and real faces); and a decreased voltage during the P200 and the N250 for the self-faces, both in the real and in the avatar conditions, as a result of familiarity processing (Tanaka et al., 2006).

The amplitude of each component was extracted for each participant in the specified time-windows (N170: 160–220 ms; P200: 250–300 ms) and analyzed via a repeated measures ANOVA with three within-subject factors: Hemisphere (left (PO7, P7), right (PO8, P8)) × Familiarity (self, familiar, other) × Virtuality (real, avatar). Importantly, only the first 100 trials of each condition are taken into account for this section to avoid the effects of adaptation (see below for the analysis of fast adaptation). The Mauchly test was run to account for significant differences on the variances. In cases where the test was positive (<0.05) we applied and noted the corresponding correction for sphericity.

In the previous section, we found the P200 component to be sensitive to both Virtuality and Familiarity. In this section, we investigate further the modulation of these effects over time. In Figure 5 we can observe the evolution of the amplitude of the P200 component for the Virtuality parameter across the whole experiment. The amplitude of the P200 component was computed in bins of 30 trials for both avatar and real faces (pooling together all the conditions) in the left hemisphere, where the effect of Virtuality was stronger (see previous section). An important observation in this Figure 5 (top panel) is the clear tendency of the voltage amplitude of the P200 component to progressively reduce over time in the avatar condition, tending to merge with the real face condition. Figure 5 (lower panel) also shows the grand average ERP for the first 50 and last 50 trials of each condition showing a clear reduction of the P200 amplitude differences in the last trials for both types of real and avatar faces.

In order to analyze these effects, we conducted a repeated measures ANOVA with factors Time (bins, 1–6) × Virtuality (real, avatar) focused on occipito-parietal left-hemisphere electrodes (PO7, P7) where the effect was observed to be larger (Figure 4). This analysis showed a significant main effect on Virtuality (F_(1,16) = 18.407, p = 0.001). Importantly and coherently with Figure 5 (top), a significant time interaction effect was found (Time × Virtuality (F_(5,80) = 2.865, p = 0.02), indicating a decay of the P200 amplitude over time for the avatar faces. Further pairwise comparisons showed that while a significant Virtuality effect was present in the first 50 trials (t = 4.037, p = 0.001, df = 16), no significant effects were observed in the last 50 trials (t = 1.169, p = 0.26, df = 16).

These findings suggest that the avatar faces are initially processed as different from the real faces, as if they were not faces but another kind of object, which is analogous to what happens with exposure to faces vs. flowers or other objects (Bentin et al., 1996; Allison et al., 1999; Keyes et al., 2010). Nevertheless, this effect is reduced over time, and both real and avatar faces eventually converge to be classified as the same class of objects: faces.

A similar analysis was run to study Familiarity effects over time at the P200 component (combining both right and left hemisphere). For this purpose, bins of 30 trials were used in the repeated measures ANOVA with factors Time (bins, 1–6) × Familiarity (self, familiar, other). A significant main effect was found for Familiarity (F_(2,32) = 5.308, p = 0.010); however, no interaction effects were observed for Time × Familiarity on the P200 component (F_(10,160) < 1). These results suggest that the effect of familiarity did not significantly decay over time in the present design. The analysis was also run with the avatar faces alone and there was no interaction of Time. This indicated that the self-recognition mechanisms are still functioning after longer exposures and both the self-avatar and the self-real faces are processed as belonging to one familiarity level, which is different to the other faces.

In the previous sections we have shown that the P200 amplitude was dependent on the Familiarity factor, and more precisely to the self-recognition in both real and avatar faces. In this section, we explore whether neurophysiological voltage could be related to higher cognitive processes such as the subjective self-identification score.

Since different participants may have different amplitudes in their components, we normalized the amplitude by calculating the voltage difference from the self-avatar to the self-real face for each participant (self-avatar—self-real) pooling together all the occipito-parietal electrodes in both hemispheres (PO7, PO8, P7, P8). A zero value coming out from that difference would indicate an equal voltage for the real picture and for the avatar, thus a greater self-identification, while more positive values would mean a lesser self-identification with the avatars, since in general a more reduced voltage in the P200 is associated with self-identification (Keyes et al., 2010).

The resulting voltage difference was then contrasted to the initial reported score of realism for the self-avatar face. In that question participants rated whether the avatar looked like themselves. A significant Spearman Correlation was found between the difference voltage (self-avatar—self-real) and the subjective score (r = −0.47, n = 17, p = 0.045; Figure 6).