Facial expressions play an important role in social life. The information is valuable for interpreting how others feel and their behavioral tendencies. Ekman (1994) classified emotional facial expressions to six basic categories (happiness, sadness, anger, disgust, fear and surprise). So far, methods such as single cell recordings, functional brain imaging and event-related potentials (ERPs) have been used to investigate brain activity involving perception, emotion, behavior, etc. Thus, studies have shown the probable neural network of emotionally salient stimuli (Eimer and Holmes, 2007). As shown by previous studies, brain activities related to emotional events, including those in the higher order sensory cortex, amygdala, orbitofrontal cortex and ventral striatum, share complex interconnected structural network. However, much more research needed to understand the brain mechanisms underlying emotion.

The recognition speed of facial expressions of emotion is very easy to obtain. Abundant data from research has revealed that the recognition speed of happy faces is faster than sad faces (Crews and Harrison, 1994; Leppänen and Hietanen, 2004; Calvo and Beltrán, 2013; Liu et al., 2013), angry faces (Billings et al., 1993; Hugdahl et al., 1993; Calvo and Beltrán, 2013), disgusted faces (Stalans and Wedding, 1985) and neutral faces (Hugdahl et al., 1993; Leppänen and Hietanen, 2004; Calvo and Beltrán, 2013; Liu et al., 2013). However, behavioral experiments can only use simple measures such as performance accuracy and reaction times (RTs). In addition, the RTs of multiple expressions classification tasks are usually 1 s or above (e.g., Calder et al., 2000; Palermo and Coltheart, 2004; Calvo and Lundqvist, 2008), which, in terms of advance chronometry, are very large time scales. For exploring more precise time processes, previous studies have explored the neural mechanism of facial expression classification using ERPs (e.g., Eimer and Holmes, 2007; Lynn and Salisbury, 2008). Some ERP research has revealed the phenomenon of positive classification advantage (PCA), which means that positive facial emotional expressions are recognized faster than negative ones (Leppänen and Hietanen, 2004). PCA is strongly linked to late perceptual processing, the differences between fearful and happy faces were shown over occipital regions as early as 80 ms post-stimuli, and those between happy and sad faces between 90 ms and 110 ms. Thus, our research focuses on this time course of PCA processing.

Many previous studies have revealed that ERP components are strongly linked to the categorization of facial emotion expression. Studies have shown that PCA is related to late components (Liu et al., 2013). However, information on the time course of facial emotion categorization has not been revealed. Brain oscillations, which could better track the activities of neurons, could be an efficient way to explore the time course of facial emotional categorizations. Oscillation activities could provide key physiologically information of brain dynamics. Furthermore, we can use neurofeedback to train different oscillation activities. Thus, we can discriminate between positive stimuli more quickly. Thus, the electroencephalogram (EEG) dynamics of face perception and facial expression have recently been analyzed through oscillation dynamics.

Event-related theta oscillations (4–7 Hz) have been reported to play an important role in cognitive processes such as memory, attention and cognition (Klimesch et al., 1997; Kahana et al., 2001; Khader et al., 2010; Sauseng et al., 2010). Balconi and Lucchiari (2006) reported enhanced frontal theta synchronization to emotional facial expressions as compared, with neutral expressions. Additionally, higher theta synchronization to fearful facial expressions than neutral expressions was observed. Similarly, Knyazev et al. (2009) reported that theta synchronization was higher in response to emotional faces (angry and happy faces) than neutral faces (Knyazev et al., 2009).

Alpha oscillations (8–15 Hz), which are pronounced due to their asymmetric effect, have been studied for many years (Davidson, 2003, 2004; Coan and Allen, 2004; Herrmann et al., 2004). Despite these obvious features on emotional processing, Güntekin and Basar (2007) found that in comparison with happy expressions, angry expressions elicited higher alpha responses at T5, P3 and O2 electrodes. Additionally, in an MEG experiment by Onoda et al., they found that event-related alpha power in the occipital region is higher in negative conditions than in other conditions (neutral and positive; Onoda et al., 2007). However, Balconi and Mazza (2009) reported that compared with neutral stimuli, positive and negative emotions trigger decreased alpha power responses. Furthermore, they also found that alpha oscillation was associated with an increase in left hemisphere activity (Balconi and Mazza, 2010). Thus, the modulation of alpha oscillations on emotional processes is still not clear.

Beta oscillations (16–30 Hz) have been thought to have a strong link with sensorimotor functions and could be reduced by voluntary movements and motor imagery (Neuper et al., 2009; Engel and Fries, 2010). Some researchers have reported enhanced beta activities in response to affective stimuli compared with neutral stimuli (Woodruff et al., 2011). Güntekin and Başar (2010) also reported higher beta activity in response to negative images than positive images in frontal, central and parietal electrodes upon presentation of IAPS images (Güntekin and Başar, 2010). In addition, Schutter et al. (2001) conducted a spontaneous EEG study and found a significant relationship in response to angry facial stimuli between asymmetry in parietal beta power and the attentional response (Schutter et al., 2001). Güntekin and Basar (2007) also reported increased beta power in response to angry facial stimuli compared with happy stimuli at F3 and Cz (Güntekin and Basar, 2007). However, Zhang et al. (2013) found greater beta oscillation activity for positive facial expressions than for negative expressions. Emotion processing mechanisms are rather complicated to reveal; thus, more studies are needed to complete our knowledge on these related brain structures.

As mentioned above, the results on the brain oscillations produced upon the processing of emotional faces have been controversial. Thus, in the present study, we will trace the time course of PCA based on brain oscillations. Our goal is to find the modulation of PCA on frequency bands within different time intervals. In prior research, Liu et al. (2013) found the N170 (150–170 ms) component, posterior N2 (250–290 ms) component and P3 (350–450 ms) component using schematic face stimuli (Liu et al., 2013). Moreover, they found that neutral faces elicited a shorter N170 latency compared with happy and sad faces. Meanwhile, they found that happy faces elicited more negative N2 activity compared with neutral and sad faces. Additionally, happy and neutral faces elicited higher P3 amplitudes and shorter P3 latencies compared with sad faces. Based on these results, we assumed that when a subject makes a decision, he/she first classifies the neutral face (time window 1: 100–200 ms) and then discriminates between happy and sad expressions (time window 2: 200–300 ms and time window 3: 100–200 ms). Thus, the time course of PCA is time window 2 and 3. We will mainly focus on these two time windows.

To test our hypothesis, we used schematic face stimuli based on Liu et al. (2013). Schematic faces allow us to control physical features as carefully as possible, to minimize influence from additional information related to facial identity (e.g., gender, race) and to exclude the confounding effects of general arousal as well as valence, per se (Boucsein et al., 2001; Eger et al., 2003; Krombholz et al., 2007; Babiloni et al., 2010). Moreover, Sagiv and Bentin (2001) have proved that even schematic faces (only made from simple line fragments) could trigger face-sensitive N170, and that this effect was not attributable to an artifact arising from facilitated recognition of a single feature (Sagiv and Bentin, 2001; Leppänen and Hietanen, 2004). In addition, schematic face stimuli are reported to be able to provide emotional stimuli; significant increase of fMRI signal can be found in the amygdala, hippocampus and prefrontal cortex in response to emotional vs. neutral schematic faces (Wright et al., 2002), indicating the feasibility for applying the schematic faces to study PCA.

Below is a summary of the results of the present research. Through facial expression classification experiments, we will discuss the phenomenon of happy face classification advantage using oscillation characteristics as well as time course. In line with previous studies, we found that face expression classification is quicker for happy expressions than for sad expressions (e.g., Crews and Harrison, 1994; Leppänen and Hietanen, 2004), and faster for neutral faces than happy faces. Moreover, happy faces elicited a higher spectral power than sad face. In accordance with other studies (Balconi and Mazza, 2009), the results showed a left hemisphere dominance, revealing the lateralization effect of positive expression classification. Although facial expressions elicited different activity on theta, alpha and beta frequencies, only beta and theta showed significant negative correlation with RT. Moreover, beta was strongly correlated with PCA in time window 1 and 3, whereas theta was correlated with time window 2. Additionally, the greater spectral power was linked to a shorter RT in all the three time windows and three frequency bands. Thus, quick responses require more brain activity, which was not obvious in the alpha bands.

In the present task, the theta oscillations showed that happy faces elicit higher power than sad faces, but the power elicited by neutral faces was higher than emotional faces (happy and sad faces). Previous studies with IAPS affective pictures as stimuli showed that emotional stimuli always elicited higher power than neutral stimuli (Balconi and Lucchiari, 2006; Zhang et al., 2013), which is in contrast with our results, but the stimuli in these experiments are not facial expressions. Studies using emotional video clips showed that power elicited by positive stimuli is higher than that by negative stimuli in the post-occipital area (Aftanas et al., 1998), which is similar to our results. Moreover, previous study using real emotional expressions suggested that happy expressions has faster results and higher accuracy than neutral expressions in discriminating tasks (Dasilva et al., 2016), which indicated the PCA.

As previously shown, alpha power showed an increase with positive stimuli in comparison with neutral stimuli after 100 ms post-stimuli, and although this difference is not significant, it is in line with previous studies that emotional stimuli elicits higher power than neutral stimuli. However, we found that in comparison with happy and neutral stimuli, power elicited by sad stimuli was the lowest. Interestingly, a study by Baumgartner et al. (2006) found that, when presenting IAPS pictures of fear, happiness and sadness, there were no differences in alpha power, but decreased alpha power was found when stimuli were emotional pictures accompanied with emotional music. For studies that found that negative stimuli elicits higher alpha responses, because negative stimuli in these studies were always angry pictures or affective pictures, the decreased alpha oscillations may be attributed to the facial classification mechanisms; however, whether real facial expressions would have same results as schematic facial expressions requires further study.

Previous studies on application of IAPS images found that negative images elicit greater beta responses compared with positive images in frontal, central and parietal electrodes (Güntekin and Başar, 2010); however, the stimuli they used were not facial expression pictures but affective pictures. A study by Zhang et al. (2013) with Chinese affective pictures as stimuli indicated that adolescents at the age of 12 exhibit more beta event-related synchronization (ERS) for positive vs. neutral stimuli. Other studies also verified that higher beta responses elicited both positive and negative stimuli than neutral stimuli (Miskovic and Schmidt, 2010; Cohen et al., 2013).

The ERP results have shown components (i.e., P1, N1, N170, P2, N2, P3) that have a strong relationship with facial emotion classification tasks. First, enhanced N170 (a negative ERP component during 140–180 ms post-stimuli at occipito-temporal electrodes), which is thought to be an indicator of inverted-face recognition, shows significant differences to fearful faces than neural faces at 160 ms post-stimuli (Holmes et al., 2005). Compared with time window 1, we concluded that face categorization was pre-attended during this time window. For posterior N2, a negativity peaking between 200 ms and 300 ms may be modulated by factors influencing visual stimuli categorization such as mutual information level, determined by gross similarity between the fragment and image in an image patch (Harel et al., 2007). In addition, as a generic name for relatively late positive component with a distribution at centro-parietal or centro-frontal midline area, P3 is considered in conjunction with facial emotion categorization (Polich, 2007). It has shown higher amplitudes and shorter latencies in response to both happy and neutral stimuli than sad stimuli, while RT displayed a significant correlation with amplitude and latency of the P3. During time windows 2 and 3, which are linked to the N2 and P3 components, emotion categorization is completed. Thus, the time-division can help us to better analyze the process of emotion categorization.

Though there were significant differences in theta, alpha and beta oscillations for different emotional stimuli (happy, neutral and sad), only theta and beta band significantly were correlated with RTs. However, the correlation between beta oscillation and RTs was as early as 100 ms after stimuli onset. That is, beta oscillation may modulate the categorization of neutral faces. In our results, the beta band had a major impact upon face discrimination. Also, the categorization occurred as early as 100 ms after stimulus onset, which has important effects for our social life. Furthermore, researchers have revealed that beta oscillation contributed to classification of known faces and unknown faces within 100–200 ms (Ozgören et al., 2005). Based on these results, we propose that in time window 1, subjects can discriminate between neutral and emotional faces, and that beta band is related to the process. This suggests that neutral face may be classified the fastest; and is modulated by beta band.

As for time windows 2 and 3, previous studies revealed that theta power was related to the P3 component (see a review Polich, 2007) and that P3 was related to PCA (Liu et al., 2013). That is, theta (time window 2) and beta (time window 3) oscillations modulate the PCA. These results are consistent with other research that focused on negative stimuli (Cohen et al., 2013). This research demonstrated that the theta band contributed to early emotion selection (200–300 ms) and that the beta band was a late response (400–600 ms). Thus, our hypothesis was verified, meaning that the theta band and beta band had a major impact on PCA.

In conclusion, our research used schematic emotional faces and focused on the time course of the recognition advantage of happy faces. The results demonstrate that the categorization process was mainly associated with beta oscillations from the beginning, while theta oscillations participated during the later period. Thus, we could detect PCA as early as 200 ms post-stimuli through theta oscillations, earlier than the other findings on time domain (usually 300 ms). Moreover, we can train theta bands in time window 2 and beta bands in time window 3, so that people can discriminate happy faces more quickly. This may have good implications for depressed patients. Therefore, advanced research on facial emotional processing is clearly needed.

TY contributed to the conception of the study. TL and XD performed the data analyses and wrote the manuscript. NM and DC contributed significantly to analysis and manuscript preparation. DC and LZ helped perform the analysis with constructive discussions. LD and CW provided data processing ideas and creative methods.

The authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.