Automatic processing of variability in multiple facial expressions

The variability of multiple facial expressions can be extracted efficiently. However, whether processing emotional variability is automatic, and whether it is impacted by types of emotions, remains unresolved. To answer these questions, we employed a passive oddball paradigm and recorded event-related potentials while participants did an attentional demanding task to detect the changes in the central fixation. A set of four faces was shown in the periphery, either displaying low (Experiment 1: SD = 21; Experiment 2: SD = 7) or high emotional variability (Experiment 1: SD = 36.12; Experiment 2: SD = 20.39), which was manipulated by changing the distance of emotional units among faces. In Experiment 1, the face set consisted of two angry and two happy faces, and the mean emotion was neutral (M = 50). In Experiment 2, all four faces were angry or happy, and the mean emotion was moderate anger (M = 25) or moderate happiness (M = 75). In Experiment 3, we additionally controlled for the range of emotional units in the set and used symmetrical distributions in both low (M = 50, SD = 35.67) and high variability conditions (M = 50, SD = 43.42). The two variability conditions had matched mean emotions and were shown with a probability of 20% (deviant) and 80% (standard) respectively in the sequence, or vice versa. When deviant stimuli are embedded within a series of standard stimuli, the appearance of the deviant would disrupt the established statistics or the regularity, leading to the observation of the vMMN (visual mismatch negativity). The vMMN is considered an index of automatic change detection, evoked by any alteration in the sequence. The results showed that in Experiment 1, faces with low emotional variability did not elicit vMMN, while those with high emotional variability elicited vMMN at both early (110-140 ms) and late (320-420 ms) time intervals. Further multivariate pattern analysis (MVPA) using all EEG channels showed that the brain could decode the standard and deviant stimuli before 100 ms under the conditions of both high and low emotional variability. In Experiment 2, when the mean emotion was angry, faces with low emotional variability elicited vMMN, whereas faces with high emotional variability elicited vMMP (visual mismatch positivity) in the time window of 320-420 ms. In contrast, when the mean emotion was happy, faces with both high and low emotional variability did not elicit significant vMMN in the 320-420 ms time window. Moreover, under all conditions of emotional variability, the standard and deviant stimuli could be successfully decoded at an early stage, but the decoding onset latency was significantly later in the low compared to the high variability condition, for both happy and angry emotions. In Experiment 3, faces with low emotional variability did not elicit vMMN, while those with high emotional variability elicited vMMP in the time window of 320-420 ms. The MVPA showed that the standard and deviant stimuli could be decoded during the early stage in both variability conditions, replicating the results of the previous experiments and excluding the potential confound of range and distribution.Taken together, we found that the variability of multiple unattended facial expressions can be perceived automatically. Moreover, there is an advantage in the automatic processing of relatively higher emotional variability, and this advantage is also influenced by the valence of emotions.