Abstract

Behavioral research has suggested a trade-off relationship between individual recognition and race categorization of own- and other-race faces, which is an important behavioral marker of face processing expertise. However, little is known about the neural mechanisms underlying this trade-off. Using functional magnetic resonance imaging (fMRI) methodology, we concurrently asked participants to recognize and categorize own- and other-race faces to examine the neural correlates of this trade-off relationship. We found that for other-race faces, the fusiform face area (FFA) and occipital face area (OFA) responded more to recognition than categorization, whereas for own-race faces, the responses were equal for the 2 tasks. The right superior temporal sulcus (STS) responses were the opposite to those of the FFA and OFA. Further, recognition enhanced the functional connectivity from the right FFA to the right STS, whereas categorization enhanced the functional connectivity from the right OFA to the right STS. The modulatory effects of these 2 couplings were negatively correlated. Our findings suggested that within the core face processing network, although recognizing and categorizing own- and other-race faces activated the same neural substrates, there existed neural trade-offs whereby their activations and functional connectivities were modulated by face race type and task demand due to one's differential processing expertise with own- and other-race faces.

Introduction

Faces are one of the most important social stimuli in our environment. The efficient and effective processing of faces plays an important role in our everyday social functioning, and difficulty in doing so can lead to debilitating social and cognitive deficits such as autism and prosopagnosia (for a review, see Lee et al. 2011). It is well established that after a protracted development, by the end of adolescence, humans become experts at processing faces (Mondloch et al. 2006; Lee et al. 2011). However, such expertise is limited only to the individuation of faces with which we encounter most frequently and have the most extensive experience (e.g., own-race faces). One of the behavioral hallmarks of such face processing expertise is characterized by a task-by-face race interaction. That is, we individually recognize own-race faces faster and more accurately than other-race faces (other-race effect [ORE]; Meissner and Brigham 2001; Sporer 2001), whereas we categorize by race other-race faces faster than own-race faces (Levin 1996; 2000; Caldara et al. 2004; Zhao and Bentin 2008; Caharel et al. 2011). These 2 effects have also been referred to as the own-race recognition advantage and the other-race categorization advantage, respectively. Both of them are robust and have been replicated numerous times using different methods, face stimuli, and participants from different race and age groups (Levin 1996, 2000; Meissner and Brigham 2001; Ge et al. 2009; Hugenberg et al. 2010; Lee et al. 2011; Anzures et al. 2013).

This task-by-face race interaction has been explained by a default processing model. It is proposed that as we become highly experienced with own-race faces, we by default recognize own-race faces at the individual level (Gauthier et al. 1999; 2000b; Tanaka 2001). In contrast, for other-race faces, we automatically pay attention to their race-diagnostic features and thereby automatically categorize them at the race level (Levin 2000; Hugenberg et al. 2010). For this reason, own-race faces are recognized better than other-race faces, whereas other-race faces are categorized by race better than own-race faces (Levin 2000; Hugenberg et al. 2010).

Further, recent behavior studies have come to suggest that these 2 race effects are not independent but interact with each other. For example, Ge et al. (2009) using a within-subject design, concurrently examined the race effects of face recognition and categorization with the same participants. They found that participants, regardless of whether they were Caucasian or Chinese, recognized own-race faces faster and more accurately than other-race faces, but categorized the race of other-race faces faster than own-race faces. Further, they found that the better an individual's recognition of own-race faces relative to other-race faces, the better the individual categorizes other-race faces relative to own-race faces. Additionally, Levin (2000) reported that the participants with a recognition advantage for own-race faces detected other-race faces more quickly than own-race faces, whereas ones without a recognition advantage for own-race faces did not present the detection advantage for other-race faces. These findings implied that the expertise at efficient recognition of own-race faces is not cost free. Rather, it is achieved at the cost of the efficiency of the categorization of these faces (Ge et al. 2009). The mirror pattern of own-race recognition advantage and other-race categorization advantage thus may reflect a trade-off between the neural processing of identity-diagnostic information and that of race-diagnostic information of own- and other-race faces (Ge et al. 2009).

No neuroimaging study has directly tested this neutral trade-off hypothesis (Ge et al. 2009). Nevertheless, some evidence from existing event-related potentials (ERP) studies seems to support this hypothesis indirectly. For example, Ito and Urland (2005) found that during a race categorization task, the N170, a face-relevant component (for a review, see Kanwisher and Yovel 2006), was more negative for own-race faces than for other-race faces. In contrast, other ERP studies reported an opposite race effect when faces were individually recognized, namely that the N170 was more negative for other-race faces than for own-race faces (Walker et al. 2008; Stahl et al. 2010; Wiese et al. 2014). Especially, Senholzi and Ito (2012) used a within-subject design similar to Ge et al. (2009) with the ERP methodology. They not only replicated Ge et al.'s (2009) behavioral findings, but also found that N170 was more negative for other-race faces than for own-race faces in the recognition task, whereas the opposite was true in the categorization task. The similar findings regarding the race effects in behavior and N170 suggest that there may be a direct relationship between the ways people process faces from different races and their brain activities. The neural trade-off hypothesis thus predicts that the responses of some cortical regions in the core face processing network (Haxby et al. 2000) to own- and other-race faces may be differentially or even oppositely modulated by task demands (i.e., recognition vs. categorization), respectively. Further, given the different functional roles of each of the face-preferential regions, the interactions between the different regions may respond differently or even in the opposite direction when recognizing and categorizing own- and other-race faces.

However, due to the low spatial resolution, the ERP methodology cannot directly test these possibilities. In contrast, the functional magnetic resonance imaging (fMRI) methodology is ideal to test the neural trade-off hypothesis because of its much higher spatial resolution than ERP. Nevertheless, in contrast to the large number of behavior and ERP studies, limited fMRI studies have examined the neural mechanism underlying each of the 2 face race effects and none has examined the 2 effects concurrently (for a review, see Ito and Bartholow 2009; Natu and O'Toole 2013). Among the few existing fMRI studies, some researchers have revealed a differential response of the right fusiform face area (FFA; Kanwisher and Yovel 2006) to faces from different races. For example, Golby et al. (2001) have found that when participants viewed and remembered faces, the response of the right FFA was greater for own-race faces than for other-race faces, whereas the activities of the left fusiform gyrus correlated with the own-race memory superiority. Kim et al. (2006) instructed participants to indicate whether the presented faces were famous or unfamiliar. They also found a greater response of the bilateral FFA to own-race faces than to other-race faces, but only for unfamiliar, not for famous faces. When faces were categorized by race, the bilateral occipital face area (OFA; Gauthier et al. 2000b) and the bilateral FFA showed greater activations to own-race faces than other-race faces (Feng et al. 2011). Additionally, using the multivoxel pattern classification method, Natu et al. (2010) also found race-related response differences in the broader ventral occipital temporal (VOT) regions. However, these fMRI studies only compared the responses of face-preferential regions between own- and other-race faces in a single task (e.g., only a face recognition task [Golby et al. 2001] or race categorization task [Feng et al. 2011]). They could not examine whether the responses of face-preferential regions (e.g., the FFA) to own-race faces or other-race faces can be differentially modulated by task demands (e.g., recognition vs. categorization) and thus failed to test the neural trade-off hypothesis directly.

A direct test of the neural trade-off hypothesis requires the use of the face recognition and categorization tasks concurrently in a within-subject design. Using such a design, an fMRI study would reveal (i) whether the responses in the face-preferential regions to own- and other-race faces will be differentially or even oppositely modulated by task demands (i.e., recognition vs. categorization) and (ii) whether the interactions between the different face-preferential regions will be modulated differentially or even oppositely when recognizing and categorizing own- and other-race faces. For example, the FFA and OFA are suggested to be generally involved in the recognition of face identity (Kanwisher and Yovel 2006). Thus, given the suggestion that other-race faces are by default categorized at the race level (Levin 2000; Hugenberg et al. 2010), can the response of the FFA and OFA to other-race faces be enhanced by an individual recognition task relative to a race categorization task? On the other hand, given the suggestion that own-race faces are automatically recognized at the individual level (Gauthier et al. 1999; 2000b; Tanaka 2001), would such areas as the superior temporal sulcus (STS; Haxby et al. 2000) show greater activation when own-race faces are categorized by race than when they are individually recognized? These questions can be readily addressed by using a within-subject design similar to Ge et al (2009) and Senholzi and Ito (2012). By contrasting activation differences of the same participants when they recognize and categorize own- and other-race faces, the effects of face race, task demand, and the interaction between these 2 factors can be directly investigated.

In addition to the examination of regional activation differences, the within-subject design could also allow for comparisons of functional connectivity differences between the key face-preferential cortical regions during face recognition and categorization and thereby would allow for a more direct testing of the neural trade-off hypothesis. Previous studies have mainly focused on the investigation of the fMRI signal changes in the face-preferential regions (e.g., FFA; Golby et al. 2001; Kim et al. 2006), but how the regions are functionally related and interact with each other for face recognition and categorization has seldom been explored (Feng et al. 2011). As generally suggested, successful face recognition and categorization depend on not only the activation of appropriate cortical regions but also on how these regions are functionally coupled between each other (Haxby et al. 2000; Fairhall and Ishai 2007; Liu et al. 2013). In fact, findings from recent fMRI studies have suggested that the functional connectivity among face-referential cortical regions might be more related to face recognition than the activities of these regions themselves. For example, recent studies have reported that some people with severely congenital or acquired prosopagnosia both still present normal face-like activation in some regions of VOT (Avidan et al. 2005; 2013; Liu et al. 2013), and the impairments in face recognition of these participants may be due to the functional dissociation between some key regions related to face recognition (Avidan et al. 2013; Liu et al. 2013). Further, a recent resting-state fMRI study, where there was not any stimulus-driven brain response, revealed that the connectivity between face-preferential regions significantly correlated with participants' behavior performance to recognize faces (Zhu et al. 2011). Thus, it is reasonable to believe that investigation of the change of functional connectivity between face-preferential regions may provide more insight into the trade-off hypothesis between recognition and categorization of own- and other-race faces. To date, no fMRI evidence exists to ascertain the validity of this intriguing hypothesis.

The present study used a within-subject 2 × 2 factorial design, namely face race (Caucasian vs. Chinese) by task type (recognition vs. categorization) to test the trade-off hypothesis between recognition and categorization of own- and other-race faces. In addition to the traditional region of interest (ROI) analysis, we used a dynamic causal modeling (DCM) analysis (Friston et al. 2003) to explore the interactions between those face-preferential regions related to the recognition or the categorization of different race faces. The DCM is a method of measuring effective connectivity and can make inferences about the directional coupling between brain regions as well as how that coupling changed as a function of experiment manipulation (Friston et al. 2003).

Materials and Methods

Participants

Twenty-four Han Chinese adults (16 males; mean age = 23.4 years, range = 1.7 years) from Beijing, PR China, where over 99.9% of the population are Han Chinese, participated the present study. All of them were healthy and right handed with normal or corrected-to-normal vision and reported not having had direct contact with other-race individuals. All subjects provided written informed consent prior to their participation in the present study, which was approved by the Human Research Protection Program of Tiantan Hospital, Beijing, China.

Stimuli

The stimuli in the present study included 64 Caucasian and 64 Chinese young adults' full-color face photos with a resolution of 640 × 480 pixels. All faces (half male and half female), which had not been seen by participants before, were upright and frontal with neutral expressions, and had their original external features (e.g., hairstyle). During the entire experiment, participant viewed all images through a mirror mounted on the top of head coil, to which the computer screen was projected.

These face images were selected in the present study for several reasons. First, they had been previously rated by both Caucasian and Chinese adults in terms of attractiveness and distinctiveness and all of them were matched on these 2 dimensions (Feng et al. 2011). Second and more importantly, those face images had been used in the Ge et al. (2009) study. Using the same face set, Ge et al. (2009) demonstrated that Chinese participants categorized the Caucasian faces better than the Chinese faces but recognized the Chinese faces better than the Caucasian faces. In contrast, Caucasian participants categorized the Chinese faces better than the Caucasian faces but recognized the Caucasian faces better than the Chinese faces. Thus, Ge et al. (2009) established the 2 race effects to be due to participants' differential experience with own- and other-race faces but not attributable to the specific faces used.

Procedure

The experiment included an event-related design task period and a block design localizer period. The task period was a 2 × 2 factorial within-participant design, namely task type (recognition vs. categorization) by face race (Caucasian vs. Chinese).

The recognition task consisted of 2 sessions, each of which adopted an old/new paradigm. Sixty-four Caucasian and 64 Chinese images were divided into 2 lists with equal numbers, which were used in different sessions, respectively. For each session, before the fMRI scanning, there was a learning phase, wherein the participants lay inside the MRI and passively viewed 16 Caucasian faces and 16 Chinese faces that were presented in a random order and repeated 3 times. Each face image was presented for 1500 ms with intervals of 500-ms of a central fixation cross between 2 adjacent face images. The participants were instructed to remember those faces as well as they possibly could. Following the learning phase closely was an fMRI scanning phase, wherein those previously learned 32 faces mixed with another 32 faces (16 Caucasian faces and 16 Chinese faces) and 30 “null events” were presented in a random order. Each face trial comprises an initial 500-ms central fixation cross followed by a 2500-ms presentation of a face stimulus, during which the participants were asked to indicate as accurately as they could whether or not they had seen that face in the learning phase by pressing a button on a response device with their left or right finger, respectively (counterbalanced across participants). Each “null event” was a 3000-ms central fixation cross and served as baseline, during which no response was needed. The fMRI scanning phase began with a 6-s central fixation cross to calm participants and ended with a 12-s central fixation cross to allow the BOLD signal elicited by the last trial to go back to baseline.

The categorization task consisted of 2 fMRI scanning sessions and included the same face images as those used in the recognition task. During each session, 32 Caucasian face images, 32 Chinese face images and 30 null events were presented in a random order. The face trials and the null events were the same as those used in the scanning phase of the recognition task, except that during the face trials participants were asked to indicate whether the currently presented face was a Caucasian face or a Chinese face. The recognition task and the categorization task were performed in a counterbalanced order across all participants.

Following the task period were 2 functional localizer sessions, each of which were comprises eight 24-s blocks of 3 categories (2 for Caucasian faces, 2 for Chinese faces, and 4 for common objects) with the intervals of 16-s of a central fixation cross between adjacent blocks. Those blocks were presented in a counterbalanced order across 2 sessions. During each block, participants viewed 16 images presented for 500 ms and followed by a 1000-ms central fixation cross and performed an occasional one-back task. The aim of the functional localizer sessions is to independently identify the traditional brain regions with preferential response to faces.

fMRI Data Acquisition

MRI data were acquired using a 3.0 Tesla MRI scanner (Siemens Trio a Tim, Germany) at Tiantan Hospital. For each of the task sessions and each of the localizer sessions, 150 and 161 whole brain T2*-weighted axial images were acquired, respectively, using standard EPI sequences (TR = 2000 ms, TE = 30 ms, FOV = 256 mm, flip angle = 90°, matrix = 64 × 64, voxel size = 4 × 4 × 4 mm3, number of slices = 32). For each participant, 3D T1-weighted structural images were acquired using a magnetization prepared rapid acquisition gradient echo sequence (voxel size: 1 × 1 × 1 mm3, FOV = 256 mm).

fMRI Data Analysis

Image data analysis included preprocessing and statistical analyses and was performed using the Statistical Parametric Mapping software (SPM8, Wellcome Trust Centre for Neuroimaging, London, UK; http://www.fil.ion.ucl.ac.uk/spm, Friston et al. 1995). During preprocessing, after spatial realignment to the first scan of the first session and slice-timing correction, all scans were co-registered to the T1-weighted structural image and normalized to the MNI (Montreal Neurological Institute) template, and then they were resampled to 2 × 2 × 2 mm3 voxels, and spatially smoothed with an isotropic 8 mm full-width-half-maximal Gaussian kernel to decrease spatial high-frequency noise and ensure the validity of inferences based on parametric tests.

After preprocessing, the task sessions and the localizer sessions were analyzed separately. For each participant, a general linear model (GLM) was constructed that included 4 regressors representing 4 experimental conditions, namely, Caucasian faces and Chinese faces in the recognition task, and Caucasian faces and Chinese faces in the categorization task. Each regressor was created by convolving a canonical hemodynamic response function with a delta function corresponding to the presentation sequence of each stimulus category. Movement parameters were used in the GLM as regressors to account for residual effects related to movement. All sessions were high-pass filtered (high-pass filter = 128 s) to remove low-frequency noise such as scanner drift (Friston et al. 1995).

For each participant, the fMRI data from the localizer sessions were analyzed with a similar GLM to that of the task sessions, except that only 2 regressors were included that represented 2 experimental conditions, namely faces (Caucasian faces and Chinese faces) and common objects.

ROI Analysis

Using the fMRI data from the localizer sessions, the traditional face-preferential cortical regions were identified with a contrast of faces minus common objects (P < 0.005, uncorrected), such as FFA, OFA, and STS. To avoid the possibility of those regions being too large to be separated from adjacent activated voxel clusters, some of them were identified with statistical thresholds of P < 0.0001 or P < 0.001. Those face-preferential regions were referred to as the ROIs in the following fMRI data analysis. It was noted that because the “face” condition included both Caucasian faces and Chinese faces, this avoids biasing the definition of ROI toward any face race (Golby et al. 2001). Within each ROI, the percent signal change (PSC) elicited by each condition relative to the baseline was calculated using MarsBar software (Brett et al. 2002). A 2 × 2 repeated two-way ANOVA, namely task type (recognition vs. categorization) by face race (Caucasian vs. Chinese), was performed on the PSC to compare the brain response to own-race faces or other-race faces between the recognition task and the categorization task.

DCM Analysis

Using the DCM analysis (Friston et al. 2003), we aimed to investigate how the response of the face-preferential regions influenced each other in the recognition task and the categorization task for own- and other-race faces. DCM takes the stimulus function that encodes the experimental manipulation as the inputs and the time course of measured fMRI responses in each brain region as the outputs. Based on the known inputs and outputs, DCM can estimate the parameters that characterize the modulatory effect reflecting the change in functionally intrinsic connections between brain regions, which is induced by the specific experimental condition (Friston et al. 2003; Mechelli et al. 2004). Thus, the DCM analysis allows us to make inferences about how the functional connectivity between face-preferential regions is modulated by own-race faces and other-race faces during the recognition task and the categorization task, respectively.

Region Selection and Time Course Extraction

As revealed by the results of the localizer sessions, 3 face-preferential regions were identified by the same contrast as that used for the ROI analysis (i.e., faces minus objects) within the right hemisphere (i.e., the right FFA, the right OFA, and the right STS), but only 2 within the left hemisphere (i.e., the left FFA and the left OFA) with the absence of the peer of the right STS in the contra-hemisphere. As a result, only the 3 face-preferential regions within the right hemisphere were included in the DCM analysis. Additionally, to ensure that as many as possible subjects had all of those 3 regions, we reduced the statistical threshold of activation comparison of face versus object to P < 0.05 (uncorrected). Such criterion was generally applied in recent studies using DCM analysis (e.g., Cardin et al. 2011; Ewbank et al. 2011; Heim et al. 2009). According to this criterion, a total of 19 participants presented all of those 3 face-preferential regions. For each participant, the volume of interest (VOI) was defined as a 4-mm-radius spherical volume centered at the maximum of each face-preferential region, and the regional response was defined as the first principal component of the time series from all voxels included in the respective VOI (Friston et al. 2003).

The DCM Model Space

Because we had few a prior predictions about the exact coupling between those VOIs in relation to the influence of different face races, it was necessary to construct a model space including all potential models and then find the optimal one using the Bayesian model selection (BMS) procedure (Stephan et al. 2009). For the purposes of the present study, we constructed the model space separately for these 2 tasks.

For the model space of the recognition task, a basic model was produced with reciprocal intrinsic connections between each pair of VOIs. The driving input included all face stimuli regardless of face race, whereas the Caucasian faces and the Chinese faces were defined as the 2 modulatory inputs, respectively. Based on the basic model, all candidate models of this model space were generated with differences in 2 factors, namely (i) the region (i.e., the FFA, the OFA or the STS) where the driving input was specified (Fig. 1) and (ii) the intrinsic connection where the modulatory input was specified (Fig. 2). Because there is not consistent evidence definitely supporting that any of those 3 VOIs send face information to the other 2 (Haxby et al. 2000; Atkinson and Adolphs 2011), each of them may be a potential region where the driving input enters. Thus, by specifying or not the driving input at each VOI, we assumed 7 (23 − 1 = 7, discount of one pattern without any driving input) possible input patterns, namely the right OFA, the right FFA, the right STS, the right OFA and the right FFA, the right OFA and the right STS, the right FFA and the right STS, and all of those 3 VOIs (Fig. 1). With respect to the modulatory inputs (i.e., Caucasian faces and Chinese faces), we either specified both of 2 modulatory inputs or none of them at a certain intrinsic connections. One of the advantages of such an arrangement is that it allows us keep the size of model space in a reasonable range. Thus, by specifying or not the 2 modulatory inputs at each intrinsic connection, 64 (26 = 64) possible modulatory patterns were generated (Fig. 2). Multiplying 64 modulatory patterns by 7 input patterns in total produced 448 candidate models for the recognition task for each participant.

Figure 1.

Seven patterns of driving input. For each of the recognition task and the categorization task, the driving input represents both Caucasian faces and Chinese faces. By specifying the driving input at different regions of the basic model, 7 (23−1 = 7, discount of one pattern without any driving input) input patterns were generated. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 1.

Seven patterns of driving input. For each of the recognition task and the categorization task, the driving input represents both Caucasian faces and Chinese faces. By specifying the driving input at different regions of the basic model, 7 (23−1 = 7, discount of one pattern without any driving input) input patterns were generated. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 2.

Sixty-four patterns of modulatory input. For each of the recognition task and the categorization task, the Caucasian faces and the Chinese faces were defined as 2 modulatory inputs, respectively. By specifying both of them or none of them at each intrinsic connection of the basic model, 64 (26 = 64) modulatory patterns were generated. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 2.

Sixty-four patterns of modulatory input. For each of the recognition task and the categorization task, the Caucasian faces and the Chinese faces were defined as 2 modulatory inputs, respectively. By specifying both of them or none of them at each intrinsic connection of the basic model, 64 (26 = 64) modulatory patterns were generated. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

With respect to the model space for the categorization task, the same 448 models were produced except that the driving input and the modulatory inputs were replaced by the corresponding stimulus functions of the categorization task.

Random-Effect BMS

After estimating all candidate models, we performed the BMS separately for the recognition task and the categorization task at group level to select the optimal modes for respective model spaces. To avoid the adverse influence of outlier participants, we adopted a random-effect (RFX) BMS at group level that is insensitive to that influence (Stephan et al. 2009; 2010). This method, using a hierarchical Beyesian model and treating the model as a random variable, allows us to estimate at a group level the exceedance probability, namely the belief that one model is more likely than any other models given group data. The exceedance probabilities are sensitive to the confidence in the posterior probability, and the BMS results indicated by them are easy to interpret because of their summation to one over all tested models, and we therefore used them to rank the superiority of each model over other models (Stephan et al. 2009).

In the present study, a two-level RFX BMS procedure was adopted, namely BMS at family level and within family level (Penny et al. 2010). To this end, corresponding to the 7 patterns of driving inputs, the 448 models of each model space were grouped in 7 families of 64 models based on the different regions of driving inputs (Fig. 1), namely family F1 with the right OFA input, family F2 with the right FFA input, family F3 with the right STS input, family F4 with the right OFA and right FFA inputs, family F5 with the right OFA and right STS inputs, family F6 with the right FFA and the right STS inputs, and family F7 with all 3 VOIs inputs. For example, for all models of family F4, the driving inputs were specified at the same VOIs, namely, the right OFA and the right FFA. It is noted that the exceedance probabilities for the family level indicate the belief that one family is more likely than any other families given group data (Penny et al. 2010). After performing the BMS at the family level, all models from the winning family were subject to a second RFX BMS (i.e., within family) to find the optimal individual model.

After the BMS procedure, the t-test was performed separately for the parameter presenting the modulatory effects at each coupling of the optimal model. The aim of these tests was to examine, across all 19 participants, whether those parameters were different from zero.

Results

Because all participants of the present study were Chinese, we will henceforth refer to Chinese faces as own-race faces and Caucasian faces as other-race faces, respectively.

Behavior Results

Table 1 summarizes the response times and the accuracies for the recognition task and the categorization task. A 2 × 2 repeated two-way ANOVA, namely task type (recognition vs. categorization) by face race (own- vs. other-race faces), was performed on the response time. An interaction effect of task type by face race was significant (F1,23 = 36.942, P < 0.001). Post hoc test between face races revealed that for the recognition task, the participants responded faster to own-race faces than to other-race faces (t23 = 2.592, P = 0.016), whereas for the categorization task the participants responded faster to other-race faces than to own-race faces (t23 = 5.523, P < 0.001). These results replicated the robust race categorization effect and other-race effect (Levin 2000; Meissner and Brigham 2001; Ge et al. 2009). The main effect of the task type was significant (F1,23 = 87.462, P < 0.001), due to the fact that, for each of own- and other-race faces, participants responded faster to the categorization task than to the recognition task (own-race faces t23 = 5.937, P < 0.001; other-race faces t23 = 10.544, P < 0.001). The main effect of the face race was also significant (F1,23 = 8.142, P = 0.009), due to the fact that the response time was significantly longer for own-race faces than for other-race faces.

Table 1

The behavior records for the recognition task and the categorization task

 Response time (ms)
 
Accuracy
 
 Caucasian faces Chinese faces Caucasian faces Chinese faces 
Recognition task 
 Mean 1028.29 993.73 0.77 0.77 
 SD 134.31 98.47 0.09 0.11 
Categorization task 
 Mean 756.71 856.22 0.98 0.96 
 SD 103.74 126.47 0.03 0.04 
 Response time (ms)
 
Accuracy
 
 Caucasian faces Chinese faces Caucasian faces Chinese faces 
Recognition task 
 Mean 1028.29 993.73 0.77 0.77 
 SD 134.31 98.47 0.09 0.11 
Categorization task 
 Mean 756.71 856.22 0.98 0.96 
 SD 103.74 126.47 0.03 0.04 

A repeated ANOVA analysis performed on the accuracies only found the significant main effect of task type (F1,23 = 119.516, P < 0.001) due to the fact that for each of own- and other-race faces, participants responded more accurately during the categorization task than during the recognition task (own-race faces t23 = 7.753, P < 0.001; other-race faces t23 = 12.005, P < 0.001). In contrast, neither the interaction effect of task type by face race (F1,23 = 1.021, P = 0.323) or the main effect of the face race (F1,23 = 2.446, P = 0.131) is significant.

ROI Analysis Results

Using the localizer sessions, 5 traditional face-preferential cortical regions, namely, the right FFA, the left FFA, the right OFA, the left OFA, and the right STS, were independently identified out of 24 participants by comparing the activation elicited by the Caucasian faces and the Chinese faces to that of common objects. We identified the right FFA from 19 participants (MNI: 41 ± 4, −49 ± 6, −19 ± 5), the left FFA from 14 participants (MNI: −41 ± 2, −54 ± 6, −21 ± 6), the right OFA from 18 participants (MNI: 40 ± 6, −76 ± 6, −13 ± 5), the left OFA from 17 participants (MNI: −38 ± 6, −79 ± 7, −14 ± 7), and the right STS from 17 participants (MNI: 47 ± 7, −49 ± 9, 10 ± 4). The loci of those face-referential regions are consistent with previous studies (e.g., Kanwisher and Yovel 2006; Fairhall and Ishai 2007; Liu et al. 2009). For convenience of description, these ROIs are referred to using these names in the present paper. However, it should be noted that it is still in debate about the response selectivity of these cortical regions (Gauthier et al. 1999; 2000a; Kanwisher and Yovel 2006).

For all ROIs, the PSC elicited by each condition of the task sessions relative to the baseline was significantly larger than baseline (Ps < 0.001) (Fig. 3).

Figure 3.

The results of analysis of region of interest. The vertical axis indicates the mean PSC across the participants. FFA, the fusiform face area; OFA, the occipital face area; STS, the superior temporal sulcus.

Figure 3.

The results of analysis of region of interest. The vertical axis indicates the mean PSC across the participants. FFA, the fusiform face area; OFA, the occipital face area; STS, the superior temporal sulcus.

A 2 × 2 repeated two-way ANOVA, namely task type (recognition vs. categorization) by face race (own- vs. other-race faces), was performed on the PSC of each ROI. In the right FFA, a significant interaction effect of task type by face race was revealed (F1,18 = 7.630, P = 0.013). However, neither the main effect of task type nor that of the face race was significant (task type: F1,18 = 3.435, P = 0.080; face race: F1,18 = 2.474, P = 0.133). The post hoc tests between tasks revealed that for other-race faces, the response of the right FFA was greater for the recognition task than for the categorization task (t18 = 2.487, P = 0.023), whereas for own-race faces there was no significant difference in response between these 2 tasks (t18 = 1.003, P = 0.329) (Fig. 3A).

Within the left FFA, similar to the right FFA, a significant interaction effect of task type by face race was observed (F1,13 = 34.271, P < 0.001). Neither the main effect of task type nor that of face race were significant (task type: F1,13 = 1.672, P = 0.218; face race: F1,13 = 1.947, P = 0.186). The post hoc test between tasks revealed that, for other-race faces, the response of the left FFA was greater for the recognition task than for the categorization task (t13 = 2.262, P = 0.041), whereas, for own-race faces, there was no significant difference in response between these 2 tasks (t13 = 0.192, P = 0.851) (Fig. 3B).

Within the right OFA, the significant interaction effect of task type by face race was revealed (F1,17 = 6.884, P = 0.018). Neither the main effect of task type nor that of the face race were significant (task type: F1, 17 = 1.346, P = 0.262; face race: F1,17 = 0.012, P = 0.914). The post hoc test between tasks revealed similar results to those of the FFAs: for other-race faces, the response of the right OFA was greater for the recognition task than for the categorization task (t17 = 2.789, P = 0.013), whereas, for own-race faces, there was no significant difference in response between these 2 tasks (t17 = 0.097, P = 0.924) (Fig. 3C).

Within the left OFA, a similar interaction effect of task type by face race to that of the right OFA was observed (F1,16 = 22.470, P < 0.001). Neither the main effect of task type nor that of the face race were significant (task type: F1,16 = 1.870, P = 0.190; face race: F1,16 = 2.229, P = 0.155). The post hoc test between tasks revealed similar results to those of the FFAs and the right OFA, namely that for other-race faces, the response of the left OFA was greater for the recognition task than for the categorization task (t16 = 2.976, P = 0.009), whereas for own-race faces, there was no significant difference in response between these 2 tasks (t16 = 0.271, P = 0.790) (Fig. 3D).

As for the PSC within the right STS, we found a significant interaction effect of task type by face race (F1,16 = 13.761, P = 0.002). Neither the main effect of task type nor that of the face race were significant (task type: F1,16 = 1.307, P = 0.270; face race: F1,16 = 2.046, P = 0.172). The post hoc test between tasks revealed opposite results to those of the FFAs and OFAs, namely that for own-race faces, the response of the right STS was greater for the categorization task than for the recognition task (t16 = 2.363, P = 0.031), whereas for other-race faces, there was not any significant difference in response between these 2 tasks (t16 = 0.205, P = 0.840) (Fig. 3E).

DCM Analysis Results

Figure 4 shows the results of RFX BMS for the recognition task (Fig. 4A) and the categorization task (Fig. 4B). For the recognition task, at the family level, the winning family was the family F1 (Fig. 4A left). The exceedance probability of this family was 0.875, surpassing those of the other families (ranging from 0.003 to 0.063). Within F1, as revealed by a second RXF BMS for all 64 models of this family, the winning model was the model M40 with a model exceedance probability of 0.159, surpassing those of the other models (ranging from 0.008 to 0.069) (Fig. 4A, middle). Figure 4A (right) shows the optimal modulatory models (M40) for the recognition task.

Figure 4.

The results of RFX BMS. For the recognition task (A), at the family level, the exceedance probability of family F1 surpassed those of the other families (A, left). Within F1, the exceedance probability of model M40 exceeded those of the other models in this family (A, middle). Thus, the M40 was identified as the optimal model for the recognition task (A, right). For the categorization task (B), at the family level, the exceedance probability of family F2 surpassed those of the other families (B, left). Within F2, the exceedance probability of model M7 exceeded those of the other models in this family (B, middle). Thus, the M7 was identified as the optimal model for the categorization task (B, right). FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 4.

The results of RFX BMS. For the recognition task (A), at the family level, the exceedance probability of family F1 surpassed those of the other families (A, left). Within F1, the exceedance probability of model M40 exceeded those of the other models in this family (A, middle). Thus, the M40 was identified as the optimal model for the recognition task (A, right). For the categorization task (B), at the family level, the exceedance probability of family F2 surpassed those of the other families (B, left). Within F2, the exceedance probability of model M7 exceeded those of the other models in this family (B, middle). Thus, the M7 was identified as the optimal model for the categorization task (B, right). FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

For the categorization task, at the family level, the winning family was the family F2 (Fig. 4B left). The exceedance probability of this family was 0.826, surpassing those of the other families (ranging from 0.001 to 0.070). Within F2, as revealed by a second RXF BMS for all 64 models of this family, the winning model was the model M7 with a model exceedance probability of 0.167, surpassing those of the other models (ranging from 0.008 to 0.059) (Fig. 4B middle). Figure 4B (right) shows the optimal modulatory model for the categorization task.

Figure 5 shows the intrinsic connectivity during the recognition task and the categorization task. As indicated by Figure 5A, when participants individually recognized the face stimuli regardless of face race, the right OFA that received the inputting information of faces excited the response of the right FFA (t18 = 3.707, P = 0.002), whereas it exerted an inhibitory influence on the response of the right STS (t18 = −3.195, P = 0.005). The right STS, in turn, also excited the response of the right FFA (t18 = 2.490, P = 0.023). In contrast, as shown in Figure 5B when participants categorized the faces by race regardless of face race, the right FFA that received the inputting information of faces excited the response of the right OFA (t18 = 3.166, P = 0.005), and the latter, in turn, exerted an inhibitory effect on the right STS (t18 = −2.555, P = 0.020).

Figure 5.

The intrinsic connectivity for the recognition task (A) and for the categorization task (B). The value indicates the cross-participants mean (SD) of the intensity of each intrinsic connection. The significant intrinsic connectivity (P < 0.05, uncorrected) and those that did not reach the significance are indicated by solid arrowed lines and dotted arrowed lines, respectively. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 5.

The intrinsic connectivity for the recognition task (A) and for the categorization task (B). The value indicates the cross-participants mean (SD) of the intensity of each intrinsic connection. The significant intrinsic connectivity (P < 0.05, uncorrected) and those that did not reach the significance are indicated by solid arrowed lines and dotted arrowed lines, respectively. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 6 shows the modulatory effect of the optimal model induced by own-race faces and other-race faces for the recognition task (Fig. 6A) and the categorization task (Fig. 6B). During the recognition task, the other-race face enhanced the reciprocal coupling between the right FFA and the right STS by 0.50 ± 0.94 (t18 = 2.306, P = 0.033) from the right FFA to the right STS and 0.42 ± 0.86 (t18 = 2.151, P = 0.045) for the reverse direction, whereas it significantly reduced the forward-feed couplings from the right OFA to the right FFA and that from the right OFA to the right STS by 0.70 ± 1.38 (t18 = −2.216, P = 0.040) and 0.70 ± 1.27 (t18 = −2.395, P = 0.028), respectively (Fig. 6A, middle-left). With respect to own-race faces, it only significantly enhanced the coupling from the right FFA to the right STS by 0.15 ± 0.27 (t18 = 2.369, P = 0.029) (Fig. 6A, middle-right).

Figure 6.

The modulatory effect of the optimal model for the recognition task (A) and for the categorization task (B). The value indicates the cross-participants mean (SD) of the modulatory effect of each intrinsic connection. The significant effects (P < 0.05, uncorrected) and those that did not reach the significance are indicated by solid arrowed lines and dotted arrowed lines, respectively. Additionally, for marginally significant effects, their P-values are also shown. The bottom line shows the correlation of modulatory effect between coupling from the right OFA to the right STS (horizontal axes) and that from the right FFA to the right STS (vertical axes) for each modulatory condition during each task. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

Figure 6.

The modulatory effect of the optimal model for the recognition task (A) and for the categorization task (B). The value indicates the cross-participants mean (SD) of the modulatory effect of each intrinsic connection. The significant effects (P < 0.05, uncorrected) and those that did not reach the significance are indicated by solid arrowed lines and dotted arrowed lines, respectively. Additionally, for marginally significant effects, their P-values are also shown. The bottom line shows the correlation of modulatory effect between coupling from the right OFA to the right STS (horizontal axes) and that from the right FFA to the right STS (vertical axes) for each modulatory condition during each task. FFA, the right fusiform face area; OFA, the right occipital face area; STS, the right superior temporal sulcus.

During the categorization task, the other-race face significantly enhanced the forward-feed coupling from the right OFA to the right STS by 0.24 ± 0.50 (t18 = 2.115, P = 0.049), whereas marginally significantly reduced the forward-feed coupling from the right FFA to the right STS by 0.42 ± 0.93 (t18 = −1.998, P = 0.061) (Fig. 6B, middle-left). In contrast, the own-race face only marginally significantly enhanced the forward-feed coupling from the right OFA to the right STS by 0.16 ± 0.35 (t18 = 2.064, P = 0.054) (Fig. 6B, middle-right).

It should be noted that, within these 2 optimal models, we are most interested in 2 of their couplings, namely the coupling from the right OFA to the right STS and that from the right FFA to the right STS, for 2 reasons. First, only these 2 couplings were included in the optimal model for the recognition task and that for the categorization task (as indicated in Fig. 6A and B, top). Further, these 2 couplings presented opposite modulatory pattern for the recognition task and the categorization task. For the categorization task, the coupling from the right OFA to the right STS was consistently enhanced by both own-race faces and other-race faces, whereas the coupling from the right FFA to the right STS was not. Moreover, the coupling from the right FFA to the right STS was even reduced by other-race faces. In contrast, for the recognition task, an opposite pattern was true, namely that the coupling from the right FFA to the STS was consistently enhanced by both own-race faces and other-race faces, whereas the coupling from the right OFA to the right STS was not. Moreover, the coupling from the right OFA to the right STS was even reduced by other-race faces. To further explore their roles in individual recognition and race categorization, we calculated the correlation of the modulatory effect between these 2 couplings for each modulatory condition (own-race faces and other-race faces) as well as each task (recognition task and categorization task). We found that for each modulatory condition and each task, the modulatory effect on coupling from the right FFA to the right STS was negatively correlated with that from the right OFA to the right STS (Ps < 0.001) (Fig. 6, bottom line).

Discussion

The aim of the present study was to examine the neural correlates underlying the trade-off between recognizing and categorizing own- and other-race faces. We specifically tested the neural trade-off hypothesis that (i) the responses of the face-preferetial regions in the core face processing network (Haxby et al. 2000) to own- and other-race faces are differentially or even oppositely modulated by task demands (i.e., recognition vs. categorization), respectively, and (ii) the interactions between the different face-preferential regions may respond differently or even in the opposite direction when recognizing and categorizing own- and other-race faces. To test this hypothesis, the present study used a 2 × 2 of task type by face race factorial design and concurrently compare the response of our brain to own-race faces and other-race faces between the categorization task and the recognition task.

Findings from ROI Analysis

We first used the ROI analysis to test whether the response pattern of the traditional face-preferential cortical regions (i.e., the FFA, OFA, and STS) to own-race faces and other-race faces during the recognition task and categorization task was consistent with the predication of the trade-off hypothesis.

The Response Patterns of the Bilateral FFA and the Bilateral OFA

One major finding of the present study was that the responses of both the FFA and OFA to own-race faces and other-race faces showed different sensitivity to task demands. When we compared the FFAs activities between tasks for own- or other-race faces, respectively, for other-race faces, the recognition task engendered greater activations than the categorization task. In contrast, for own-race faces, both tasks elicited similar levels of activation. In other words, both the right and the left FFA are sensitive to task demand only for other-race faces but not for own-race faces. Like the FFAs' results, for other-race faces, both OFAs responded significantly more during the recognition task than during the categorization task. However, for own-race faces, both tasks elicited similar levels of activation.

Because the same face images were used in the 2 tasks, the differential responses to the other-race faces between tasks in the FFAs and OFAs could only be attributed to the difference in task demand. Recent converging fMRI evidence have demonstrated that not only the FFA but also the OFA play an important role in recognition of face identity and both regions may contain neural populations tuned specifically to face identity (e.g., Schiltz et al. 2006; Schiltz and Rossion 2006; Rhodes et al. 2009; for a review, see Atkinson and Adolphs 2011). Given the specific roles of the FFA and the OFA in the processing of face identity, their differential sensitivity to the manipulation of task demand for own- and other-race faces is consistent with the default processing model described above (Levin 2000; Tanaka 2001; Hugenberg et al. 2010). According to the default processing model (Levin 2000; Tanaka 2001; Hugenberg et al. 2010), we have expertise with own-race faces and whereby automatically encode own-race faces at the individual level regardless of task demands. Therefore, no additional cognitive efforts are needed to extract identity information for own-race faces during the recognition task relative to the categorization task. Consistent with this, we found that the FFAs' responses or the OFAs' responses to own-race faces were not significantly different for those 2 tasks. However, for other-race faces, we have not enough expertise to drive an automatic individual recognition of them. As a result, additional cognitive efforts are needed to extract identity-diagnostic information during the recognition task, therefore resulting in stronger response of the FFAs and OFAs to other-race faces during the recognition task than the categorization task. This was also perhaps why in the present study it took longer response times to recognize individual other-race faces than own-race faces.

As revealed by behavior results, for each of own- and other-race faces, participants responded faster and more accurately to the categorization task than to the recognition task, suggesting more performance difficulty for the recognition task than for the categorization task. Such increase in task difficulty may be due to that the recognition task included more cognitive processing, for example, the retrieval from memory of the viewed faces in the learning phase. Thus, some may argue that the enhanced responses of the FFA and the OFA during the recognition task relative to the categorization task might be due to this difficulty difference between the 2 tasks. However, if this was the case, for each of own- and other-race faces, the FFA and the OFA should have presented enhanced responses during the recognition task relative to during the categorization task. However, we found that for each of the FFA and the OFA, its response to other-race faces was greater during the recognition task than during the categorization task, whereas its response to own-race faces is equal for these 2 tasks. Additionally, the right STS was even activated more by own-race faces for the categorization task than for the recognition task. Thus, the task-relevant increase in the responses of the FFA and the OFA to other-race faces was not attributable to the difficulty difference between the recognition task and the categorization task.

The Response Pattern of the Right STS

In contrast to the findings of the FFAs and OFAs, the right STS showed the opposite pattern: the right STS responded equally to other-race faces for the 2 tasks, but significantly more to own-race faces for the categorization task than the recognition task. Our findings suggest that the right STS plays a different role from the FFAs and the OFAs in processing own- and other-race faces. According to the default processing model, the right STS may be involved in processing face race information as opposed to individual identity information. As our expertise with own-race faces drives an automatic individuation of own-race faces at the individual level (Gauthier et al. 2000a; Tanaka 2001), it interferes with the race categorization of own-race faces (Ge et al. 2009; Hugenberg et al. 2010; Senholzi and Ito 2012). As a result, additional cognitive resources are needed to categorize own-race faces. The significantly increased activation to own-race faces during the categorization task relative to the recognition task in the right STS might reflect the recruitment of additional neural resources needed to extract race-diagnostic information. However, for other-race faces, because the default mode is already at the race level, no additional cognitive processes are needed and hence there were no significant differences in activation in the right STS between the categorization task and recognition tasks for other-race faces. This is also perhaps why in the present study it took longer response times for the race categorization of own-race faces than that of other-race faces.

It is generally accepted that the right STS is an important part of the face processing network (Haxby et al. 2000). It has been thought to be involved mainly in the processing of changeable aspects of the face such as facial motion, emotion, and gaze (Haxby et al. 2000; Winston et al. 2004; Pelphrey et al. 2005; Engell and Haxby 2007; Pitcher et al. 2011). However, recent studies demonstrated that the right STS may be involved in the processing of not only the changeable facial aspects but also the invariant facial features (Yovel and Kanwisher 2005; Liu et al. 2009). Our findings suggest that one of the functions of the right STS for processing the invariant aspects of the face might be the processing of its racial information.

Taken together, the findings from traditional ROI analysis revealed that the response of each face-preferential region within core face processing network to own- and other-race faces was differentially modulated by task demands. Further, the FFAs and the OFAs presented opposite response pattern to that of the right STS. These findings were consistent with the predication of the trade-off hypothesis and demonstrated that the responses of the face-preferential regions were not only dependent of face race but also of the task demand.

Findings From the DCM Analysis

By analysis of the effective functional connectivity between the face-preferential regions, DCM allowed us to test which coupling was modulated by face races and task demands, and the relationship between these functional couplings.

The Intrinsic Coupling for the Recognition Task and the Categorization Task

One major findings from the DCM analysis is that although, as indicated by the ROI analysis, there were not main effects of fMRI response of the task type within each of the face-preferential regions (i.e., the right OFA, the right FFA, and the right STS), theses 3 regions were intrinsically connected in different ways for the recognition task and the categorization task, respectively. For the recognition task, the visual face information, regardless of the face race, entered into the right OFA, and then propagated to the right FFA through a forward-feed intrinsic connection, consistent with the hierarchical model of face recognition suggested by the previous studies (Haxby et al. 2000; Fairhall and Ishai 2007). Further, the response of the right FFA was also influenced by the response of the right STS through an excitatory intrinsic connection between them, supporting the core role of the right FFA in face recognition. In contrast, for the categorization task, the visual face information directly entered the right FFA, and then propagated to the right OFA through a backward-feed intrinsic connection, suggesting that such backward-feed coupling may be related to the categorization of face race.

The Modulatory Effect for own-race Face and Other-race Face During the Recognition Task and the Categorization Task

What we are most interesting is how the intrinsic connections are modulated by own- and other-race faces during the recognition task and the recognition task. We found that for the recognition task, the connectivity from the right FFA to the right STS was positively enhanced by own-race faces and other-race faces, respectively, whereas the connectivity from the right OFA to the right STS was negatively modulated by other-race faces but not own-race faces. The results for the categorization task were exactly the opposite: the connectivity from the right OFA to the right STS was positively enhanced by own-race faces and other-race faces, respectively, whereas the connectivity from the right FFA to the right STS was negatively modulated by other-race faces but not own-race faces.

Recently, some fMRI studies have investigated the role of the functional connectivity between face-relevant regions in face processing. For example, fMRI studies have found a face-specific functional connectivity between the right FFA and the right STS in the resting state (Turk-Browne et al. 2010). Further, there is also converging evidence suggesting that not only the right STS but also the functional connectivity between the right STS and the right FFA are sensitive to changes in face identity (Winston et al. 2004; Fox et al. 2009; Baseler et al. 2014). As for the coupling from the right OFA to the right STS, in the present study, it was enhanced by race categorization of own-race faces as well as other-race faces. Additionally, it is well established that the right OFA was involved in the processing facial features (Haxby et al. 2000), and the right STS had been demonstrated by our ROI analysis results to be related to the race categorization of faces. This evidence implied that this feedforward coupling might serve to regulate the processing of the invariant aspects of the faces as race. Thus, the coupling from the right FFA to the right STS and that from the right OFA to the right STS might play important roles in recognition of face identity and categorization of face races, respectively. If this is the case, these findings implied that whether the right STS plays a role in individual recognition or race categorization of faces depends on the modulatory pattern of connectivity from other regions to it. In other words, when it is positively influenced by the right FFA but negatively influenced by the right OFA, it may be engaged during the individuation of faces. In contrast, when it is positively influenced by the right OFA but negatively influenced by the right FFA, it may play a different role in race categorization.

Further, we also found that for each modulatory condition (i.e., own- or other-race faces) of each task (the recognition or the categorization task), the modulatory effect of the coupling from the right FFA to the right STS and that of the coupling from the right OFA to the right STS were negatively correlated. In other words, the more positive the modulatory effect for one of the couplings (e.g., from the right FFA to the right STS), the more negative the modulatory effect for the other coupling (e.g., from the right OFA to the right STS).

The negative correlation between the modulatory effect of these 2 couplings as well as the mirror modulatory patterns of these 2 couplings between the recognition task and the categorization task suggested a neural trade-off between individual recognition and race categorization of faces. Thus, when recognizing the identity of a face, especially an other-race face, people have to overcome the interference from race categorization and shift their attention from the race-diagnostic features to identity-diagnostic information of that face. To support such an operation, the neural activities related to the recognition of face identity must be enhanced, whereas those related to the categorization of face race may need to be suppressed. In contrast, when categorizing the race of a face, especially an own-race face, the opposite behavior and neural operations would result. Perhaps for these reasons, our functional connectivity findings revealed an opposite connectivity modulatory patterns.

In addition to the couplings from the right FFA to the right STS and that from the right OFA to the right STS, the one from the right STS to the right FFA and that from the right OFA to right FFA are positively and negatively modulated, respectively, when other-race faces are individually recognized. As suggested by previous studies, the coupling from the right OFA to the right FFA is generally suggested to be involved in face recognition (Haxby et al. 2000). However, in the present study, it was decreased for the recognition of other-race faces. Our findings suggested that for other-race faces, the recognition of their identity may be dependent on a reciprocal coupling between the right FFA and the right STS rather than a feedforward influence from the right OFA to the right FFA.

Taken together, the findings from the DCM analyses revealed that the functional connectivity between face-preferential regions presented opposite modulatory patterns for the recognition task and the categorization task, respectively. These 2 modulatory patterns were negatively correlated, suggesting a neural trade-off between the processing of identity-diagnostic information and that of race-diagnostic information of faces.

In summary, using fMRI methodology, we concurrently asked participants to recognize and categorize own- and other-race faces to examine the neural correlates of the trade-off relationship between individual recognition of own- and other-race faces and the categorization of them by race. We found that for other-race faces, the FFAs and OFAs responded more to the recognition task than the categorization task, whereas for own-race faces, the responses were equal for the 2 tasks. The right STS's responses were the opposite to those of the FFAs and the OFAs. Further, recognition enhanced the functional connectivity from the right FFA to the right STS, whereas categorization enhanced the functional connectivity from the right OFA to the right STS. The modulatory effects of these 2 couplings were negatively correlated. Our findings suggested that within the core face processing network, although recognizing and categorizing own- and other-race faces activated the same neural substrates, there existed neural trade-offs whereby their activations and functional connectivities were modulated by face race type and task demand due to one's differential processing expertise with own- and other-race faces.

Funding

This work was supported by the National Basic Research Program of China (973 Program) under Grant 2011CB707700, the National Natural Science Foundation of China under Grant No. 81227901, 61231004, 61375110, 30970771, 60910006, 31028010, 30970769, 81000640, the Fundamental Research Funds for the Central Universities (2013JBZ014, 2011JBM226) and NIH (R01HD046526 and R01HD060595).

Notes

Conflict of Interest: None declared.

References

Anzures
G
Quinn
PC
Pascalis
O
Slater
AM
Tanaka
JW
Lee
K
.
2013
.
Developmental origins of the other-race effect
.
Curr Dir Psychol Sci
 .
22
:
173
178
.
Atkinson
AP
Adolphs
R
.
2011
.
The neuropsychology of face perception: beyond simple dissociations and functional selectivity
.
Philos T R Soc B
 .
366
:
1726
1738
.
Avidan
G
Hasson
U
Malach
R
Behrmann
M
.
2005
.
Detailed exploration of face-related processing in congenital prosopagnosia: II. Functional neuroimaging findings
.
J Cognitive Neurosci
 .
17
:
1150
1167
.
Avidan
G
Tanzer
M
Hadj-Bouziane
F
Liu
N
Ungerleider
LG
Behrmann
M
.
2013
.
Selective dissociation between core and extended regions of the face processing network in congenital prosopagnosia
.
Cereb Cortex
 .
(in press)
.
Baseler
HA
Harris
RJ
Young
AW
Andrews
TJ
.
2014
.
Neural Responses to Expression and Gaze in the Posterior Superior Temporal Sulcus Interact with Facial Identity
.
Cereb Cortex
 .
24
:
737
744
.
Brett
M
Anton
J
Valabregue
R
Poline
J
.
2002
.
Region of interest analysis using an SPM toolbox [abstract]
.
Paper presented at the 8th International Conference on Functional Mapping of the Human Brain
,
Sendai, Japan
.
Available on CD-ROM in NeuroImage. 16(2)
.
Caharel
S
Montalan
B
Fromager
E
Bernard
C
Lalonde
R
Mohamed
R
.
2011
.
Other-race and inversion effects during the structural encoding stage of face processing in a race categorization task: An event-related brain potential study
.
Int J Psychophysiol
 .
79
:
266
271
.
Caldara
R
Rossion
B
Bover
P
Hauert
CA
.
2004
.
Event-related potentials and time course of the ‘other-race’ face classification advantage
.
NeuroReport
 .
15
:
905
910
.
Cardin
V
Friston
KJ
Zeki
S
.
2011
.
Top-down modulations in the visual form pathway revealed with dynamic causal modeling
.
Cereb Cortex
 .
21
:
550
562
.
Engell
AD
Haxby
JV
.
2007
.
Facial expression and gaze-direction in human superior temporal sulcus
.
Neuropsychologia
 .
45
:
3234
3241
.
Ewbank
MP
Lawson
RP
Henson
RN
Rowe
JB
Passamonti
L
Calder
AJ
.
2011
.
Changes in “top-down” connectivity underlie repetition suppression in the ventral visual pathway
.
J Neurosci
 .
31
:
5635
5642
.
Fairhall
SL
Ishai
A
.
2007
.
Effective connectivity within the distributed cortical network for face perception
.
Cereb Cortex
 .
17
:
2400
2406
.
Feng
L
Liu
J
Wang
Z
Li
J
Li
L
Ge
L
Tian
J
Lee
K
.
2011
.
The other face of the other-race effect: An fMRI investigation of the other-race face categorization advantage
.
Neuropsychologia
 .
49
:
3739
3749
.
Fox
CJ
Moon
SY
Iaria
G
Barton
JJ
.
2009
.
The correlates of subjective perception of identity and expression in the face network: an fMRI adaptation study
.
Neuroimage
 .
44
:
569
580
.
Friston
KJ
Harrison
L
Penny
W
.
2003
.
Dynamic causal modeling
.
NeuroImage
 .
19
:
1273
1302
.
Friston
KJ
Holmes
AP
Worsley
KJ
Poline
JP
Frith
CD
Frackowiak
RSJ
.
1995
.
Statistical parametric maps in functional imaging: a general linear approach
.
Hum Brain Mapp
 .
2
:
189
210
.
Gauthier
I
Skudlarski
P
Gore
JC
Anderson
AW
.
2000a
.
Expertise for cars and birds recruits brain areas involved in face recognition
.
Nat Neurosci
 .
3
:
191
197
.
Gauthier
I
Tarr
MJ
Anderson
AW
Skudlarski
P
Gore
JC
.
1999
.
Activation of the middle fusiform ‘face area’ increases with expertise in recognizing novel objects
.
Nat Neurosci
 .
2
:
568
573
.
Gauthier
I
Tarr
MJ
Moylan
J
Skudlarski
P
Gore
JC
Anderson
AW
.
2000b
.
The fusiform “face area” is part of a network that processes faces at the individual level
.
J Cogn Neurosci
 .
12
:
495
504
.
Ge
L
Zhang
H
Wang
Z
Quinn
PC
Pascalis
O
Kelly
D
Slater
A
Tian
J
Lee
K
.
2009
.
Two faces of the other-race effect: recognition and categorisation of Caucasian and Chinese faces
.
Perception
 .
38
:
1199
1210
.
Golby
AJ
Gabrieli
JDE
Chiao
JY
Eberhardt
JL
.
2001
.
Differential responses in the fusiform region to same-race and other-race faces
.
Nat Neurosci
 .
4
:
845
850
.
Haxby
JV
Hoffman
EA
Gobbini
MI
.
2000
.
The distributed human neural system for face perception
.
Trends Cogn Sci
 .
4
:
223
233
.
Heim
S
Eickhoff
SB
Ischebeck
AK
Friederici
AD
Stephan
KE
Amunts
K
.
2009
.
Effective connectivity of the left BA 44, BA 45, and inferior temporal gyrus during lexical and phonological decisions identified with DCM
.
Hum Brain Mapp
 .
30
:
392
402
.
Hugenberg
K
Young
SG
Bernstein
MJ
Sacco
DF
.
2010
.
The categorization-individuation model: An integrative account of the other-race recognition deficit
.
Psychol Rev
 .
117
:
1168
1187
.
Ito
TA
Bartholow
BD
.
2009
.
The neural correlates of race
.
Trends Cogn Sci
 .
13
:
524
531
.
Ito
TA
Urland
GR
.
2005
.
The influence of proessing objectives on the perception of faces: an ERP study of race and gender perception
.
Cogn Affect Behav Ne
 .
5
:
21
36
.
Kanwisher
N
Yovel
G
.
2006
.
The fusiform face area: a cortical region specialized for the perception of faces
.
Philos Trans R Soc B
 .
361
:
2109
2128
.
Kim
J
Yoon
H
Kim
B
Jeun
S
Jung
S
Choe
B
.
2006
.
Racial distinction of the unknown facial identity recognition mechanism by event-related fMRI
.
Neurosci Lett
 .
397
:
279
284
.
Lee
K
Anzures
G
Quinn
PC
Pascalis
O
Slater
A
.
2011
.
Development of face processing expertise
. In:
Calder
AJ
Rhodes
G
Johnson
MH
Haxby
JV
, editors.
Handbook of face perception
 .
Oxford (UK)
:
Blackwell Publishing
. p.
753
778
.
Levin
DT
.
1996
.
Classifying faces by race: the structure of face categories
.
J Exp Psychol Learn
 .
22
:
1364
1382
.
Levin
DT
.
2000
.
Race as a visual feature: using visual search and perceptual discrimination tasks to understand face categories and the cross-race recognition deficit
.
J Exp Psychol Gen
 .
129
:
559
574
.
Liu
J
Harris
A
Kanwisher
N
.
2009
.
The perception of face parts and configurations: an fMRI study
.
J Cogn Neurosci
 .
22
:
203
211
.
Liu
J
Wang
M
Shi
X
Feng
L
Li
L
Thacker
JM
Tian
J
Shi
D
Lee
K
.
2013
.
Neural correlates of covert face processing: fMRI evidence from a prosopagnosic patient
.
Cereb Cortex
 
(in press)
.
Mechelli
A
Price
CJ
Friston
KJ
Ishai
A
.
2004
.
Where bottom-up meets top-down: neuronal interactions during perception and imagery
.
Cereb Cortex
 .
14
:
1256
1265
.
Meissner
CA
Brigham
JC
.
2001
.
Thirty years of investigating the own-race bias in memory for faces: a meta-analytic review
.
Psychol Public Pol L
 .
7
:
3
35
.
Mondloch
CJ
Maurer
D
Ahola
S
.
2006
.
Becoming a face expert
.
Psychol Sci
 .
17
:
930
934
.
Natu
V
O'Toole
AJ
.
2013
.
Neural perspectives on the other-race effect
.
Vis Cogn.
 
21
:
1081
1095
.
Natu
V
Raboy
D
O'Toole
AJ
.
2010
.
Neural correlates of own- and other-race face perception: Spatial and temporal response differences
.
NeuroImage
 .
54
:
2547
2555
.
Pelphrey
KA
Morris
JP
Michelich
CR
Allison
T
McCarthy
G
.
2005
.
Functional anatomy of biological motion perception in posterior temporal cortex: an fMRI study of eye, mouth and hand movements
.
Cereb Cortex
 .
15
:
1866
1876
.
Penny
WD
Stephan
KE
Daunizeau
J
Rosa
MJ
Friston
KJ
Schofield
TM
Leff
AP
.
2010
.
Comparing families of dynamic causal models
.
PLOS Comput Biol
 .
6
(3)
:
e1000709
.
Pitcher
D
Dilks
DD
Saxe
RR
Triantafyllou
C
Kanwisher
N
.
2011
.
Differential selectivity for dynamic versus static information in face-selective cortical regions
.
NeuroImage
 .
56
:
2356
2363
.
Rhodes
G
Michie
PT
Hughes
ME
Byatt
G
.
2009
.
The fusiform face area and occipital face area show sensitivity to spatial relations in faces
.
Eur J Neurosci
 .
30
:
721
733
.
Schiltz
C
Rossion
B
.
2006
.
Faces are represented holistically in the human occipito-temporal cortex
.
NeuroImage
 .
32
:
1385
1394
.
Schiltz
C
Sorger
B
Caldara
R
Ahmed
F
Mayer
E
Goebel
R
Rossion
B
.
2006
.
Impaired face discrimination in acquired prosopagnosia is associated with abnormal response to individual faces in the right middle fusiform gyrus
.
Cereb Cortex
 .
16
:
574
586
.
Senholzi
KB
Ito
TA
.
2012
.
Structural face encoding: how task affects the N170's sensitivity to race
.
Soc Cogn Affect Neur
 .
8
(8)
:
937
942
.
Sporer
SL
.
2001
.
Recognizing faces of other ethnic groups an integration of theories
.
Psychol Public Pol L
 .
7
:
36
97
.
Stahl
J
Wiese
H
Schweinberger
SR
.
2010
.
Learning task affects ERP-correlates of the own-race bias, but not recognition memory
.
Neuropsychologia
 .
48
:
2027
2040
.
Stephan
KE
Penny
WD
Daunizeau
J
Moran
RJ
Friston
KJ
.
2009
.
Bayesian model selection for group studies
.
NeuroImage
 .
46
:
1004
1017
.
Stephan
KE
Penny
WD
Moran
RJ
den Ouden
HE
Daunizeau
J
Friston
KJ
.
2010
.
Ten simple rules for dynamic causal modeling
.
Neuroimage
 .
49
:
3099
3109
.
Tanaka
JW
.
2001
.
The entry point of face recognition: evidence for face expertise
.
J Exp Psychol Gen
 .
130
:
534
543
.
Turk-Browne
NB
Norman-Haignere
SV
McCarthy
G
.
2010
.
Face-specific resting functional connectivity between the fusiform gyrus and posterior superior temporal sulcus
.
Front Hum Neurosci
 .
4
:
176
.
Walker
PM
Silvert
L
Hewstone
M
Nobre
AC
.
2008
.
Social contact and other-race face processing in the human brain
.
Soc Cogn Affect Neur
 .
3
:
16
25
.
Wiese
H
Kaufmann
JM
Schweinberger
SR
.
2014
.
The neural signature of the own-race bias: evidence from event-related potentials
.
Cereb Cortex.
 
24
:
826
835
.
Winston
JS
Henson
RNA
Fine-Goulden
MR
Dolan
RJ
.
2004
.
fMRI-adaptation reveals dissociable neural representations of identity and expression in face perception
.
J Neurophysiol
 .
92
:
1830
1839
.
Yovel
G
Kanwisher
N
.
2005
.
The neural basis of the behavioral face-inversion effect
.
Neuron
 .
15
:
2256
2262
.
Zhao
L
Bentin
S
.
2008
.
Own- and other-race categorization of faces by race, gender, and age
.
Psychon B Rev
 .
15
:
1093
1099
.
Zhu
Q
Zhang
J
Luo
YLL
Dilks
DD
Liu
J
.
2011
.
Resting-state neural activity across face-selective cortical regions is behaviorally relevant
.
J Neurosci
 .
31
:
10323
10330
.