Distinct neural-behavioral correspondence within face processing and attention networks for the composite face effect

This is a self-archived version of an original article. This version may differ from the original in pagination and typographic details. Author(s): Title: Year: Version: Copyright: Rights: Rights url: Please cite the original version: CC BY-NC-ND 4.0 https://creativecommons.org/licenses/by-nc-nd/4.0/ Distinct neural-behavioral correspondence within face processing and attention networks for the composite face effect © 2021 The Authors. Published by Elsevier Inc. Published version Chen, Changming; Lou, Yixue; Li, Hong; Yuan, Jiajin; Yang, Jiemin; Winskel, Heather; Qin, Shaozheng Chen, C., Lou, Y., Li, H., Yuan, J., Yang, J., Winskel, H., & Qin, S. (2022). Distinct neural-behavioral correspondence within face processing and attention networks for the composite face effect. Neuroimage, 246, Article 118756. https://doi.org/10.1016/j.neuroimage.2021.118756 2022 NeuroImage 246 (2022) 118756 Contents lists available at ScienceDirect NeuroImage journal homepage: www.elsevier.com/locate/neuroimage Distinct neural-behavioral correspondence within face processing and attention networks for the composite face effect Changming Chen a , Yixue Lou b , c , Hong Li d , e , ∗∗ , Jiajin Yuan e , ∗ , Jiemin Yang e , Heather Winskel f , Shaozheng Qin g , h a School of Education, Chongqing Normal University, Chongqing 401331, China b Department of Psychology, Faculty of Education and Psychology, University of Jyvaskyla, Jyväskylä 40014, Finland c Faculty of Psychology, Southwest University, Chongqing 400715, China d School of Psychology, South China Normal University, Guangzhou 510631, China e Institute of Brain and Psychological Sciences, Sichuan Normal University, Chengdu 610066, China f Psychology, James Cook University, Singapore Campus, 387380, Singapore g State Key Laboratory of Cognitive Neuroscience and Learning and IDG/McGovern Institute for Brain Research, Beijing Normal University, Beijing 100875, China h Chinese Institute for Brain Research, Beijing, China a r t i c l e i n f o Keywords: Face recognition Functional magnetic resonance imaging Representational similarity analysis Insula Fusiform gyrus Composite face effect a b s t r a c t The composite face effect (CFE) is recognized as a hallmark for holistic face processing, but our knowledge remains sparse about its cognitive and neural loci. Using functional magnetic resonance imaging with independent localizer and complete composite face task, we here investigated its neural-behavioral correspondence within face processing and attention networks. Complementing classical comparisons, we adopted a dimensional reduction approach to explore the core cognitive constructs of the behavioral CFE measurement. Our univariate analyses found an alignment effect in regions associated with both the extended face processing network and attention networks. Further representational similarity analyses based on the Euclidian distances among all experimental conditions were used to identify cortical regions with reliable neural-behavioral correspondences. Multidimensional scaling and hierarchical clustering analyses for neural-behavioral correspondence data revealed two principal components underlying the behavioral CFE effect, which fit best to the neural responses in the bilateral insula and medial frontal gyrus. These findings highlight the distinct neurocognitive contributions of both face processing and attentional networks to the behavioral CFE outcome, which bridge the gaps between face recognition and attentional control models. 1. Introduction The concept of holistic processing is central to face recognition research ( Richler et al., 2011 ). It is generally believed that faces are perceived as indecomposable wholes rather than as a collection of separate parts ( Richler et al., 2012 ; Rossion, 2013 ; Tanaka & Farah, 1993 ). Holistic processing can be characterized by multiple different phenomena ( Maurer et al., 2002 ; Richler et al., 2012 ), such as the “face inversion effect ”where people recognize upright faces better than inverted faces ( Yin, 1969 ), and the “whole-part effect" where the recognition of a face part is better when presented in a whole face compared to when presented in isolation ( Tanaka and Farah, 1993 ). In the current study we investigate the composite face effect (CFE) which has gained growing interest and provides one of the most powerful pieces of evidence for holistic processing ( Hole, 1994 ; Rossion, 2013 ; Young et al., 1987 ). ∗ Institute of Brain and Psychological Sciences, Sichuan Normal University, Chengdu 610066, China ∗∗ School of Psychology, South China Normal University, Guangzhou 510631, China E-mail addresses: [email protected] (H. Li), [email protected] , [email protected] (J. Yuan). Unlike the whole-part effect which emphasizes the facilitation effect of holistic processing on immediate and long-term face memories, the CFE effect emphasizes holistic attention ( Richler et al., 2012 ; Tanaka and Simonyi, 2016 ). It describes a difficulty in selective processing of a target face part (e.g. the top half) while ignoring the remaining part of a face (e.g. the bottom half), probably due to the interference of the latter ( Richler et al., 2012 ). The CFE effect is named after the composite face paradigm which has been refined over the past three decades ( Gauthier and Bukach, 2007 ; Hole, 1994 ; Young et al., 1987 ). In the original version of this paradigm the top halves are either identical or different but the bottom halves are always different, and the CFE is indexed by an alignment effect in trials with identical tops ( Rossion, 2013 ; Young et al., 1987 ): participants are prone to errors when matching two identical target halves that are aligned with different irrelevant halves, but show high performance https://doi.org/10.1016/j.neuroimage.2021.118756 . Received 11 August 2021; Received in revised form 14 November 2021; Accepted 22 November 2021 Available online 27 November 2021. 1053-8119/© 2021 The Authors. Published by Elsevier Inc. This is an open access article under the CC BY-NC-ND license ( http://creativecommons.org/licenses/by-nc-nd/4.0/ ) C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Fig. 1.. Design and example stimuli of the complete composite face paradigm (CCP). Participants are asked to match the top halves of two composite faces indicated by the brackets. In the congruent condition the top and bottom halves are either both identical or both different, whereas in the incongruent condition they are identical at one location but different in the other location. The top and bottom halves are either aligned or misaligned. Portrait rights were consented by the persons depicted. when the target and irrelevant halves are misaligned. To minimize the confound of response bias, an additional manipulation of “congruency ” is later included in research articles ( “complete composite paradigm ”, CCP, see Fig. 1 , Gauthier and Bukach, 2007 ). In the “congruent ”trials the top and bottom halves are either both identical or both different, whereas in the “incongruent ”trials the two parts are identical in one location (e.g., the top) but different in the other (e.g., the bottom). Researchers have repeatedly observed the modulation of congruency and spatial alignment: the response is facilitated (higher discrimination index and shorter reaction time) in the “congruent ”than the “incongruent ”trials when faces are aligned, while this congruency effect is reduced or disappears when faces are misaligned (Cheung et al., 2008; Richler and Gauthier, 2014). Since its introduction, the CFE effect has emerged as one of the triedand-true measures of holistic processing ( Richler and Gauthier, 2013 ; Rossion, 2013 ; Young et al., 1987 ). It has been intensively investigated, stimulating ongoing interests and discussions about the nature of holistic processing, such as its domain-specificity ( Chua and Gauthier, 2020 ; Gauthier and Tarr, 2002 ; Rossion, 2013 ; Bauser et al., 2011 ; Zhao et al., 2015 , 2016 ), development ( Grand et al., 2004 ; Ventura et al., 2018 ), association with face recognition ability (significant in DeGutis et al., 2013 ; Richler et al., 2011 ; Wang et al., 2012 ; but null in Konar et al., 2010 ; Rezlescu et al., 2017 ; Zhao et al., 2014 ) and its link to perceptual expertise ( Gauthier and Bukach, 2007 ; Gauthier and Tarr, 1997 ; Zhao et al., 2015 ). It has also become a paradigm with much potential in delineating emotion recognition ( Calder et al., 2000 ; Xie and Zhang, 2016 ), social cognition ( Chen et al., 2018 ; Todorov et al., 2010 ), modelling of face processing ( Dailey and Cottrell, 1999 ; Fitousi, 2013 ), aberrant face perception in clinical populations such as autism ( Gauthier et al., 2009 ) and prosopagnosia (significant in Avidan et al., 2011 ; Busigny et al., 2010 ; Liu and Behrmann, 2014 ; Ramon et al., 2010 , but null in Biotti et al., 2017 ; Le Grand et al., 2006 ; Susilo et al., 2010 ), contributing substantially to our understanding of face and object recognition ( Murphy et al., 2017 ). But along with the popularity of the CFE effect is its unresolved neurocognitive operations, especially how it is affected by perceptual and attention processes. According to the template hypothesis ( Jacques and Rossion, 2009 ; Rossion, 2013 ; Tanaka and Farah, 1993 ), face recognition relies heavily on simultaneous perceptual integration of separate features which fuses facial parts into a quintessential whole (Gestalt), it is realized by a template-matching process guided by an inherently holistic internal representation (page 145, Rossion, 2013 ). In some articles the CFE is argued as a perceptual illusion (page 140, Rossion, 2013 ) unique in upright faces ( McKone et al., 2013 ), and roles of factors like object-based attention or decisional components are rejected as “they are unable to account for the CFE effect ”(page 191, Rossion, 2013 ; McKone et al., 2013 ). In terms of its dynamics, the template hypothesis proposes that the CFE is initiated by perpetual integration from early onset ( Jacques and Rossion, 2009 ; Rossion, 2013 ) implemented in face-sensitive areas including the fusiform and occipital face areas along the ventral occipitotemporal cortex ( Schiltz et al., 2010 ; Schiltz and Rossion, 2006 ). Consistent with this, fMRI studies with block-design ( Schiltz and Rossion, 2006 ) and event-related design ( Schiltz et al., 2010 ) find neural adaptations in the bilateral middle fusiform gyrus and inferior occipital gyrus when identical top halves were aligned with different bottom halves than the condition where identical tops were aligned with identical bottoms, but not in misaligned or inverted faces. Electrophysiological studies observed an increase in the N170 component in the occipto-temporal cortex following changes of identity in the top and bottom halves ( Kuefner et al., 2010 ; Von Der Heide et al., 2018 ). The within-network connectivity in the rFFA correlates negatively with the composite-face effect, and the functional connectivity between rFFA and bilateral posterior superior temporal sulcus is also negatively linked to the behavioral CFE effect (Li et al., 2019). The CFE effect is reduced following transcranial direct current stimulation (tDCS) over the occipitotemporal cortex ( Yang et al., 2014 ), or lesions in the anterior region of the right temporal lobe ( Busigny et al., 2014 ). However, it is worth noting that face processing in the OFA is more part-based ( Liu et al., 2010 ; Zhang et al., 2015 ) and there is coexistence of both part-based and whole-based representation of the FFA ( Harris and Aguirre, 2008 , 2010 ; Liu et al., 2010 ). Differences in neural response between aligned and misaligned composite faces are also reported in regions including the lateral occipital complex, parahippocampal gyrus ( Foster et al., 2015 ) and even the prefrontal cortex ( Schiltz et al., 2010 ). Mixed pictures are suggested about the neural correlates of the CFE effect, within and outside the face network. In parallel, the CFE effect is also observed in non-face objects such as greebles ( Gauthier and Tarr, 1997 ; Gauthier et al., 1998 ), visual words ( Wong et al., 2012 ) or even inverted faces ( Susilo et al., 2013 ). The expertise hypothesis proposes that holistic face processing arises from a learned strategy of attending to all parts due to substantial experience in face individuation ( Gauthier and Bukach, 2007 ; Gauthier and Tarr, 1997 ; Gauthier et al., 1998 ). Specifically, because “individuation in everyday life encourages attention to all parts ”( Chua et al., 2014 ), the attentional strategy to full face becomes automatic and cannot be easily “turned off” after extensive experience in face individuation. The CFE effect then arises as a failure of selective attention because of the “inflexibility in attentional weightings on facial parts ”(page 4, Richler et al., 2012 ). According to this view, holistic processing can be observed both in faces and non-face objects given a massive amount of individuation experience ( Gauthier and Tarr, 1997 ; Wong et al., 2012 ). In line with this, individuation experience attending to the whole face particularly the diagnostic features lead to holistic processing ( Chua et al., 2014 ; Chua et al., 2015 ) and the level of experience in individuating objects of a category determines the extent of holistic processing in new objects of that category ( Chua and Gauthier, 2020 ). fMRI studies also show neural correlates of expertise in FFA. There is, for instance, higher FFA response to objects of expertise in experts than novices ( Gauthier et al., 2000 ), as well as correlation between FFA response to the expertise objects and behavioral indices of holistic processing of those objects ( Gauthier and Tarr, 2002 ; Ross et al., 2018 ; Wong et al., 2009 ). Nevertheless, the CFE effect is also observed in non-face object without expertise such as line patterns ( Zhao et al., 2015 ). The shape-based account proposes there might be two routes for the holistic processing ( Foster et al., 2021 ; Zhao et al., 2016 ): it can arise either via a top-down 2 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 expertise route or via a bottom-up route relying merely on Gestalt information. This account suggests holistic processing is tuned to facial shape information as “shape information alone is sufficient to elicit the composite face effect ” regardless of its origins from innate face template or automatized attention ( Zhao et al., 2016 ). Recruiting the complete composite paradigm, Foster et al. (2021) observed higher activities to holistic versus part-based processing in areas within the face-responsive network (e.g. FFA2) and areas outside such as the lateral occipital complex, transverse occipital sulcus, retrosplenial cortex and the parahippocampal place area. An additional hypothesis ( Fitousi, 2015 , 2016 ) also agrees with the domain-generality of the CFE effect, but surmises CFE reflects a general attentional strategy akin to object-based attention. Different from the aforementioned “integrating ” hypotheses, Fitousi and colleagues propose the CFE effect does not necessitate the integration of information in the top and bottom face halves, but could be due to differences in the relative discriminability between them ( Fitousi, 2016 ; Fitousi and Algom, 2006 ), or due to decisional factors across conditions ( Fitousi, 2016 ). The reduction in the congruity composite effect with misalignment is likely due a similar disruption of objecthood ( Fitousi, 2015 ) as in object-based attention all constituent features or parts can be activated once an object has been selected. Computational modelling based on multidimensional signal detection theory suggests the CFE is affected by not only perceptual, but also attentional and decisional components ( Richler et al., 2009 ; Richler et al., 2008 ). Analysis of theoretical ex-Gaussian parameters of RT distributions ( Fitousi, 2020 ) also reveals that the CFE effect is generated by pure changes in the exponential component of the ex-Gaussian distribution, suggesting the involvement of attentional and working memory processes in the composite face effect regardless of partial and complete designs. Other study also revealed behavioral evidence for violations of perceptual separability, decisional separability, and believed that “holistic processing could very well be a product of both perceptual and decisional influences even when overt response is not required ”( Von Der Heide et al., 2018 ). The involvement of these loci can be further supported by neuroimaging observation that the change of identity of the top of the composite face is associated with a signal increase at the later decisional P3b component and the lateralized readiness potential (LRP) ( Kuefner et al., 2010 ). Taken together, these studies generally agree on the involvement of a perceptual component in the CFE effect, but hold different views about the role of other cognitive processes such as attentional, decisional or working memory. It is theoretically interesting to admit and elaborate on the contribution of perceptual components to the CFE effect, but increasing evidence shows clearly that the composite face effect is to some degree modulated by other sources of influence ( Liu et al., 2020 ; Murphy et al., 2017 ; Von Der Heide et al., 2018 ). Therefore, to get a fair understanding of the CFE effect, we believe it is of equal importance to adopt a more inclusive approach, targeting the debated cognitive processes directly and weighing their neural substantiation in the composite face task. However, till now the majority of existent studies emphasize the role of face regions, though a few studies have inspected regions beyond the face network their interests are still restricted to regions within the ventral or dorsal perceptual streams ( Foster et al., 2021 ), the neural operations beyond remains also largely unknown. Meanwhile, methodologically most of the previous studies examined the neural correlates of the CFE via top-down approaches ( Foster et al., 2021 ; Schiltz et al., 2010 ; Schiltz and Rossion, 2006 ), using univariate statistical parametric mapping methods contrasting conditions critical for a theoretical inference. While informative, these approaches are subject to limitations in the ability to explore representation of multidimensional data ( Popal et al., 2020 ) and may overlook valuable information in condition-rich designs like the complete composite paradigm. Sophisticated multivariate, model-free techniques such as representational similarity analysis (RSA, Kriegeskorte et al., 2006 ; Kriegeskorte et al., 2008 ) can make the enrichment. The RSA uses distance measures such as response-pattern dissimilarities (1 - correlation) or the Euclidean distance among the experimental conditions ( Kriegeskorte et al., 2008 ) to characterize the high-order representational space. Combining with second-level reduction techniques such as multidimensional scaling, it can reveal the relationship among multiple experimental conditions/manipulations and visualize the internal structure of mental operations ( Kriegeskorte et al., 2008 ). It can be used to fit neural activities in cortical regions to behavioral profiles or computational models ( Arbula et al., 2021 ; Qin et al., 2014 ). During the past decade, the RSA has been widely used to explore the response similarity between species ( Kriegeskorte et al., 2008 ), stages of cognitive processing ( Tzagarakis et al., 2009 ), cortical regions and modalities of brainactivity measurement ( Benjaminsson et al., 2010 ; Kriegeskorte et al., 2008 ), being particularly beneficial in condition-rich experimental design ( Kriegeskorte et al., 2008 ). Given these theoretical and methodological concerns, the current study investigated the roles of the face-processing network and the attentional network in the CFE, with especial interest in which networks demonstrate satisfactory neural-behavioral fitting. We conducted magnetic resonance imaging when participants received the complete composite task, and independent face localizer scanning to define these two networks. The attentional interference network was defined by the Eriksen Flanker task —a popular paradigm measuring attentional interference ( Eriksen and Eriksen, 1974 ), and the face processing network was identified by the one-back recognition task involving face and non-face stimuli. Complementing the conventional univariate analysis, we recruited the representational similarity analysis to assess the relationship between behavioral and neural activity patterns. Based on the Euclidian distances among all eight conditions, we conducted multidimensional scaling analysis and hierarchical clustering to explore the principal components underlying the experimental conditions and to visualize the respectively behavioral and neural representational (dis)similarity structure. We then explored behavioral-neural correspondence via ROI-based analyses and whole-brain searchlight analysis, based on Spearman’srank correlation between the dissimilarity matrices. The results are expected to shed light on the implementation of perceptual and attentional processes in the CFE effect, and contribute to the neurocognitive framework of face processing. 2. Methods The study was approved by the Brain Imaging Center Institutional Review Board of Southwest University, China. Informed consent was obtained in written form from all participants prior to the study. 2.1. Participants Twenty-two undergraduate or graduate students (13 females, 21.451.37 years old) from a university in mainland China participated in all the tasks and MRI scanning as detailed below (also see supplementary materials). Satisfactory statistical power was expected in the current design as a previous study reported a sample size of 15 participants could detect acceptable effect size both in behavioral and fMRI measurements using the same complete composite face paradigm ( Foster et al., 2021 ). None of the participants had a history of neurological or psychiatric illness, or reported difficulty in recognizing faces and everyday objects. They were right-handed and had normal or corrected-to-normal visual ability. They scored 59.36 ± 8.91 (range: 41–72, total score:72) in the Chinese version of the Cambridge Face Memory Test ( McKone et al., 2012 ) and 45.86 ± 7.09 (range: 35–65, total score:72) in the Cambridge Car Memory Test ( Dennett et al., 2012 ). 2.2. The experimental design The fMRI scanning was carried out in two sessions. The first session consisted of high-resolution structural imaging, face-network localizer and Eriksen Flanker localizer tasks, as well as rest-state and DTI scanning 3 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 (not reported here). The second session was carried out later within three days, where participants received 4 runs of functional imaging when performing the complete composite face task. Stimuli used in the second session were not seen in the first session. 2.3. Tasks and materials Face-network Functional Localizer Task: The localizer scanning consisted of 3 runs. Each run lasted for 5 min 12 secs and presented the following five categories in blocks: faces, houses, words, cars and scrambled words. Participants were asked to perform a one-back task detecting stimulus repetition, where the occurrence of repetition was no more than twice in each block. All the blocks were 12-s long and sandwiched by 8-s fixations. In each block 10 stimuli were presented sequentially: each was presented for 600 ms and followed by a 600 ms fixation. In total, there were 9 blocks for each category. The order of the categories was counterbalanced across runs and participants. Eriksen Flanker Task: In this task, participants were asked to judge the orientation of the central arrow, which was flanked by two stimuli on each side. The flanker stimuli were either arrows orienting to the same direction (congruent trials), or arrows orienting to the opposite direction to the central arrows (incongruent trials) or stars (neutral trials). These three trial types were interspersed randomly across two runs. Each run lasted for 6 min 20 s, and contained 40 trials for each trial type. The trials from the three conditions showed up in a pseudo-random way such that there were no more than 3 consecutive trials with the same response. In each trial, the arrow image was presented for 500 ms, and followed by a fixation for either 1500 ms or 3500 ms. Complete Composite Face Task: A total of 16 grayscale young adult Chinese faces (8 females) were used to create the composite faces. All the raw faces had neutral expressions, were the same age, grayscale, and equalized in physical properties like luminance and contrast. External features such as hair, eye glasses, and earrings were removed. There were no distinguishing features (moles or scars) that could facilitate feature-based face recognition. To generate the composites, all faces were horizontally divided into two halves in the middle, the top half of each face was aligned with the bottom half of two others of the same gender. Composites with unsatisfactory face configuration were discarded according to the following criteria: misalignment in the contours of the top and lower halves, misalignment between the nose from the center of the mouth, irregular ratio of upper lip height to mouth width ( < 0.3 or > 0.7) based on a norm of Chinese population ( Wang et al., 2019 ). Thirty-two aligned composite faces were constructed following this procedure. Misaligned faces were then generated from these composites by shifting the bottom half 70 pixels rightward and inserting a horizontal red line (3 pixels) at the conjunction of two halves. We also added a white bracket over the head of all pictures to signify the target location. All stimuli were presented at the center of the screen at a distance of 100 cm. All the aligned composites were 130 ×170 pixels (visual angle: 1.76 o ×2.29 o ), and the misaligned composites were 200 ×170 pixels (visual angle: 2.70 o ×2.29 o ). Participants were asked to match the top halves of the two composite faces (see Fig. 1 b). There were 4 runs of scanning. Each run lasted for 9 min 56 s and contained 144 trials, with 18 trials in each of the 8 conditions (aligned/misaligned ×congruent/incongruent ×same/different response). At the beginning of each run, a white bracket showed up in the upper half of the screen to notify participants that the top halves should be compared. Each trial began with a study face (300 ms), then a mask (120 ms), the cue bracket (280 ms), and finally a test face (300 ms). Between each two trials there was a fixation for either 1000, 3000 or 5000 ms. 2.4. Imaging parameters Whole-brain anatomical and functional images were collected using a 12-channel head coil on a 3.0-T Siemens Trio MRI scanner (Siemens Medical, Erlangen, Germany). The functional images were collected with a T2 ∗ -sensitive Echo Planar Imaging sequence (Sequence = epfid2d1.64), with equal parameters for all three tasks: TR/TE = 2000 ms/30 ms, flip angle = 90°, number of slices = 32, slice thickness = 3.0 mm, slice gap = 1.00 mm, matrix = 64 ×64, voxel size = 3.4375 ×3.4375 ×3 mm 3 . The whole-brain High-resolution T1-weighted anatomical images were collected using a magnetizationprepared rapid gradient echo (MPRAGE) sequence, with parameters as below: TR/TE/TI = 1900 ms / 2.52 ms / 900 ms, flip angle = 9°, number of slices = 176, slice thickness = 1.0 mm, matrix = 256 ×256, voxel size = 1 ×1 ×1 mm 3 . 2.5. Image preprocessing and GLM analysis Data were analyzed using the Analysis of Functional NeuroImages software (AFNI, http://afni.nimh.nih.gov/afni/ ) and in-house Matlab codes (Mathworks, MA). Functional images in all three tasks were preprocessed following the same pipeline. The first four volumes of each run were discarded for stabilization of magnetic resonance signal. The remaining were corrected for slice acquisition timing, and spatially realigned to correct for head movement. Data for all the tasks were aligned to the first volume of the face network localizer, and then scaled against the mean value of each run. In the univariate analysis, the scaled time courses were smoothed by a Gaussian Kernel of 6 mm FWHM. Deconvolution analyses were then conducted to estimate the regression beta values for conditions in each task based on GLM theory. The basis function for the 5 categories in face network localizer task was BOX(1,12), and the basic function for predictors in other two tasks was GAMA. Six parameters of head movement were also included in the regression model. The beta value maps were then normalized to the Talairach space and resampled into resolution of 3 ×3 ×3 mm 3 . 2.6. ROI analysis Definition of regions of interest ROIs in the face network were identified by contrasting BOLD response for faces versus houses and cars, each ROI was extracted at the individual level at the height threshold of p < 0.05. ROIs in the Eriksen Flanker Task were regions with higher BOLD response to the “incongruent ”than the “congruent ”trials based on the results of group analysis, corrected by FDR method at p < 0.05. Similar to other studies using the complete composite face paradigm, the trials were grouped into four categories at the combination of alignment and congruency, collapsing top-same and top-different trials in each category to obtain the discrimination index and mean reaction time. However, as the top-same and top-different trials contribute differentially to the behavioral composite face effect and share distinct neural representations, it is recommended that separate analysis of these two types of trials may help clarify their roles in measuring holistic processing ( Foster et al., 2021 ). Therefore, we grouped the trials into 8 conditions at the combination of the three factors: alignment (aligned, misaligned) ×congruency (congruent, incongruent) ×response type (same, different) and first ran univariate analyses on the BOLD signal estimated by general linear model. Additionally, we ran a representational similarity analysis (RSA) to quantitatively explore how well neural responses in regions of face processing and attentional interference networks fit the behavioral results. In each region, we constructed the activity-pattern dissimilarity matrix of the 8 conditions based on the Euclidean distance, and visualize the similarity structure of neural activities of the conditions combining multidimensional scaling with hierarchical clustering. The reference behavior-based dissimilarity matrix was estimated in the same way based on the inverse efficiency score (IES = RT/ACC), which takes the joint contribution of accuracy and reaction time into consideration. The behavioral-neural correspondence in each region of interest was then estimated by Spearman’s rank correlation between the upper triangular elements of the two matrices. The neural-behavioral correspondence in a region was assessed by a random permutation simulation (10000 times), 4 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Fig. 2.. The behavioral results in the complete composite paradigm (CCP). There is (A) significant main effect of congruency, main effect of alignment and the interaction between alignment and congruency in the discrimination index, (B) significant main effects of congruency and alignment in reaction time. (C) The accuracy and (D) reaction time of the 8 conditions in the CCP task. (E) The three major clusters among the 8 conditions revealed by hierarchical clustering analysis, and (F) the distribution of these clusters along the two major dimensions from multidimensional scaling analysis. (legend labels denote the 8 conditions, ACS: aligned congruent same, ACD: aligned congruent different, AIS: aligned incongruent same, AID: aligned incongruent different, MCS: misaligned congruent same, MCD: misaligned congruent different, MIS: misaligned incongruent same, MID: misaligned incongruent different). where elements in the neural dissimilarity matrix were randomized and correlated to the behavioral matrix in each permutation. The correspondence was significant if the output index was among the lowest 50 in the permutation results (equal to q ≤ 0.05 by Bonferroni correction). 2.7. Whole-brain searchlight analysis To explore regions outside the above-identified ROIs that also demonstrated significant behavioral-neural correspondence, we conducted whole-brain searchlight analyses with the afore-mentioned multidimensional scaling and Spearman’s-rank correlation. Specifically, the beta values of the 8 conditions were estimated by GLM analysis based on normalized but spatially unsmoothed images and extracted from a searchlight with radius of 3 voxels ( Anzellotti et al., 2014 ). The voxels by conditions matrix within the searchlight was then centralized to compute the dissimilarity matrix based on the Euclidean distance, which was then correlated with the reference behavioral dissimilarity matrix by Spearman’s Rank correlation. The correlation parameters images were then spatially smoothed with a kernel of FWHM of 6 mm. Significant clusters were determined based on Monte-Carlo simulation with the 3dClustSim function in AFNI, where the spherical autocorrelation –acf function parameters were determined by 3dFWHMx. 2.8. Data and code availability statement Behavioral data, scripts for data analysis is available via https://osf.io/fsqaj/ , fMRI Data can be accessed from the first author on request upon a formal data usage agreement. 3. Results 3.1. Behavioral results We first analyzed the effect of alignment and congruency on discrimination sensitivity following the conventional procedure. Discrimination index was estimated individually for each of the alignments by congruency combinations by the following formula: ( Stanislaw and Todorov, 1999 ): 𝐴 ′= 0 . 5 + 𝑠𝑖𝑔𝑛 ( 𝐻 − 𝐹 ) ( 𝐻− 𝐹 ) 2 + |𝐻− 𝐹 | 4 𝑚𝑎𝑥 ( 𝐻,𝐹 )−4 𝐻𝐹 , where H indicated hit rate and F indicated false alarm rate. Two-way ANOVA analysis (see Fig. 2 A) revealed a significant main effect of alignment ( F (1,21) = 23.794, p < 0.001, 𝜂2 𝑝 = 0 . 531 ) and congruency ( F (1,21) = 22.317, p < 0.001, 𝜂2 𝑝 = 0 . 515 ), the discrimination sensitivity was generally higher in the misaligned trials than aligned trials, and higher in congruent trials than incongruent trials. There was also a significant alignment by congruency interaction ( F (1,21) = 28.142, p < 0.001, 𝜂2 𝑝 = 0 . 573 ). Further analysis showed that the discrimination index was significantly lower in the incongruent condition than in the congruent condition when faces were aligned ( F (1,21) = 26.833, p < 0.001, 𝜂2 𝑝 = 0 . 561 ), but with no congruency effect when faces were misaligned ( F (1,21) = 0.165, p = 0.689 , 𝜂2 𝑝 = 0 . 008 ). As for reaction time (see Fig. 2 B), there was a significant main effect of alignment ( F (1,21) = 46.992, p < 0.001, 𝜂2 𝑝 = 0 . 691 ), main effect of congruency ( F (1,21) = 85.635, p < 0.001, 𝜂2 𝑝 = 0 . 803 ), and significant alignment by congruency interaction ( F (1,21) = 25.189, p < 0.001, 𝜂2 𝑝 = 0 . 545 ). The reaction time was generally slower in the aligned than the misaligned condition, and slower in the incongruent trials than congruent trials. It was slower in the “incongruent ”than the “congruent ” condition both when the face parts were aligned ( p < 0.001) and misaligned ( p < 0.001), the interaction however, suggested the congruency effect in misaligned trials was smaller than that in the aligned trials. To characterize in-detail the response pattern, we conducted a repeated-measures ANOVA on accuracy and reaction time of all 8 conditions. The analysis revealed a significant three-way interaction in accuracy ( F (1,21) = 25.292, p < 0.001, 𝜂2 𝑝 = 0 . 546 , Figure 2 C) and reaction time ( F (1,21) = 59.462, p < 0.001, 𝜂2 𝑝 = 0 . 739 , Fig. 2 D). A simple effect analysis on accuracy found a congruency by response interaction when faces were aligned ( F (1,21) = 18.191, p < 0.001, 𝜂2 𝑝 = 0 . 464 ) but not misaligned ( F (1,21) = 2.522, p = 0.127). When faces were aligned, response accuracy in congruent trials was higher than the incongruent trials regardless of the response type ( “same ”, F (1,21) = 23.412, p < 5 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Fig. 3.. The ROIs identified by independent localizer scanning. 0.001, 𝜂2 𝑝 = 0 . 527 , “different ”, F (1,21) = 5.29, p = 0.032, 𝜂2 𝑝 = 0 . 201 ), but the significant interaction indicated a higher congruency effect in aligned over misaligned faces. As for the three-way interaction in RT, there was an interaction between congruency and response when faces were aligned ( F (1,21) = 87.672, p < 0.001, 𝜂2 𝑝 = 0 . 807 ), but not when they were misaligned ( F (1,21) = 0.878, p = 0.359). In the aligned trials, RT in the incongruent condition was longer than that in the congruent condition when the correct response was “the same ” ( F (1,21) = 111.908, p < 0.001, 𝜂2 𝑝 = 0 . 842 ), but there was no congruency effect when the correct response was “different ”( F (1,21) = 1.098, p = 0.307). In both accuracy and reaction time, there was also a main effect of alignment (ACC: F (1,21) = 23.193, p < 0.001, 𝜂2 𝑝 = 0 . 525 , RT: F (1,21) = 53.245, p < 0.001, 𝜂2 𝑝 = 0 . 717 ), as well as alignment by congruency interaction (ACC, F (1,21) = 36.184, p < 0.001, 𝜂2 𝑝 = 0 . 633 , RT: F (1,21) = 42.916, p < 0.001, 𝜂2 𝑝 = 0 . 671 ), alignment by response interaction (ACC: F (1,21) = 21.546, p < 0.001, 𝜂2 𝑝 = 0 . 506 , RT: F (1,21) = 34.735, p < 0.001, 𝜂2 𝑝 = 0 . 623 ), and congruency by response interaction (ACC, F (1,21) = 11.851, p < 0.001, 𝜂2 𝑝 = 0 . 361 , RT: F (1,21) = 45.359, p < 0.001, 𝜂2 𝑝 = 0 . 684 ). Next, we employed hierarchical clustering and multidimensional scaling analysis to explore the structure among the 8 conditions. These dimension reduction techniques enable us to explore relationships among conditions via a data-driven approach. Dissimilarity matrices for these two analyses were built based on the inverse efficiency score (RT/ACC) in each condition, which has the advantage of combining the effect of both indices (Bruyer and Brysbaert, 2011; Townsend and Ashby, 1978). We also estimated the percentage of variance explained by each component based on the eigenvalues available in multidimensional scaling analysis. The results showed that the first dimension accounted for 95.18% and the second dimension accounted for 3.93% of the total variance in the inverse efficiency score. It is worth noting that similar results were also obtained in analysis on accuracy, with 85.50%, 10.53% of the variance explained by the first and second components respectively (see supplementary materials). As shown in Fig. 2 E, we identified three clusters from behavioral response of the 8 conditions. The first cluster contained only the AIS condition (alignment: aligned, congruency: incongruent, response: same), which was distant from all other conditions. The second included the ACD (aligned, congruent, different), AID (aligned, incongruent, different), MCD (misaligned, congruent, different), and MID (misaligned, incongruent, different) conditions, where the target halves were all “different ”. The third cluster contained the MCS (misaligned, congruent, same), MIS (misaligned, incongruent, same) and ACS (aligned, congruent, same) conditions where the target halves were all “the same ”. We could see in Fig. 2 F the first dimension generated by multidimensional scaling analysis was largely characterized by how far away the AIS condition was from all other 7 conditions, while the second dimension reflected largely the type of response, with all conditions of the response “different ”at one end and most conditions of the response “same ”at the other end. Similar dimensions and clusters were replicated in analyses on accuracy (see supplementary materials). 3.2. FMRI results in localizer scanning Face processing network: Five major regions, including the right Fusiform Face Area (rFFA), the left Fusiform Face Area (lFFA), the right Superior Temporal Sulcus (rSTS), the left Superior Temporal Sulcus (lSTS) and the right Occipital Face Area (rOFA) were identified individually for the participants (see Fig. 3 ). The average coordinates, number of voxels, and ratio of positive participants are available in Table 1 , detailed information including the coordinates, number of voxels, threshold of each participant can be seen in the supplementary materials (see also in Tables S2–8). Attentional Network : Group analysis revealed 5 major clusters after FDR correction ( Table 2 , Fig. 3 ), which included the right anterior insula, left anterior insula, dorsal medial prefrontal cortex (mPFC), left inferior parietal lobe (lIPL) and left precentral gyrus (lPCG). These regions replicated the classical results found in previous studies (Casey et al., 2000). 3.3. Univariate analysis in the composite face task Next, we ran a 2 by 2 by 2 ANOVA analysis to explore the impact of the three factors (alignment, congruency, response type) on BOLD response in each ROI ( Fig. 4 , Figures S5-1-10). In regions of the face network, there were significant main effects of alignment in left FFA ( F (1,21) = 15.217, p = 0.001, 𝜂2 𝑝 = 0 . 420 ), right STS ( F (1,21) = 12.224, p = 0.002, 𝜂2 𝑝 = 0 . 368 ), both survived Bonferroni-correction for N = 10 ROIs. There was main effect of alignment in right FFA ( F (1,20) = 9.553, p = 0.006, 𝜂2 𝑝 = 0 . 323 ), right OFA ( F (1,19) = 5.500, p = 0.030, 𝜂2 𝑝 = 0 . 224 ) which did not survive the Bonferroni-correction (N = 10 ROIs). In the first four regions, the BOLD response was generally higher in the aligned trials than the misaligned trials. There was also a main effect of response type in the right OFA ( F (1,19) = 10.877, p = 0.004, 𝜂2 𝑝 = 0 . 364 ), with higher beta value in trials with the response “different ”than those with the response “the same ”. No other effects were observed within the face network, p > 0.100. As for regions in the attention network, there were main effects of alignment in the left anterior insula ( F (1,21) = 11.671, p = 0.003, 𝜂2 𝑝 = 0 . 357 , surviving Bonferroni-correction for N = 10 ROIs), right anterior insula ( F (1,21) = 7.884, p = 0.011, 𝜂2 𝑝 = 0 . 273 ) and mPFG ( F (1,21) = 5.086, p = 0.035, 𝜂2 𝑝 = 0 . 195 ) which did not survive Bonferroni-correction, with higher response in the aligned trials over the misaligned trials. There were also main effects of congruency in the right anterior insula ( F (1,21) = 18.622, p < 0.001, 𝜂2 𝑝 = 0 . 470 ), mPFG ( F (1,21) = 17.593, p < 0.001, 𝜂2 𝑝 = 0 . 456 ). In all three areas, BOLD response in the incongruent trials was higher than that in the congruent trials. There were main effects of response type in the right anterior in6 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Table 1. The characteristics of face processing areas. ROI X Y Z voxels ratio rFFA 40.5 ± 3.55 -45.929 ± 5.418 -15.5 ± 3.674 60.222 ± 41.347 21/22 lFFA -39.545 ± 4.572 -47.591 ± 8.047 -15.909 ± 7.249 40.182 ± 29.936 22/22 lSTS -47.318 ± 8.342 -57.545 ± 11.086 8.773 ± 7.685 90.455 ± 72.072 22/22 rOFA 37.95 ± 6.329 -74.85 ± 7.631 -11.15 ± 5.284 46.9 ± 61.068 20/22 rSTS 52.364 ± 7.882 -52.091 ± 10.404 6.591 ± 6.248 126.857 ± 89.337 22/22 Fig. 4.. Beta values of the conditions in each ROI. Table 2. The characteristics of the attentional network. ROI X Y Z voxels rInsula 34.5 16.5 14.5 54 lInsula -31.5 16.5 14.5 39 lPCG -25.5 -7.5 53.5 117 lIPL -46.5 -37.5 44.5 134 mPFG -10.5 4.5 50.5 55 sula ( F (1,21) = 13.328, p = 0.001, 𝜂2 𝑝 = 0 . 388 ), and left anterior insula ( F (1,21) = 12.343, p = 0.002, 𝜂2 𝑝 = 0 . 370 ) where neural activity in the “same ”trials was higher than that in the “different ” trials. There were also alignment by congruency interactions in the right anterior insula ( F (1,21) = 8.410, p = 0.009, 𝜂2 𝑝 = 0 . 286 ) and mPFG ( F (1,21) = 4.312, p = 0.05, 𝜂2 𝑝 = 0 . 170 ) which did not survive Bonferroni-correction. In both areas, the incongruent trials elicited relatively higher response than the congruent trials (right anterior insula, p < 0.001, mPFG, p = 0.001) when face halves were aligned, but not when faces were misaligned (right anterior insula, p = 0.409, mPFG, p = 0.301). We also observed interactions between alignment and response type in right anterior insula ( F (1,21) = 9.137, p = 0.006, 𝜂2 𝑝 = 0 . 303 ), left anterior insula ( F (1,21) = 6.026, p = 0.023, 𝜂2 𝑝 = 0 . 223 ), and mPFG ( F (1,21) = 5.025, p = 0.036, 𝜂2 𝑝 = 0 . 193 ) which did not survive Bonferroni-correction. In all three areas, trials with the response “same ”induced higher response than those with the response “different ”when face halves were aligned (right anterior insula, p < 0.001, mPFG, p < 0.001, the left anterior insula, p < 0.001), but not when face halves were misaligned (right anterior insula, p = 0.077, mPFG, p = 0.67, left anterior insula, p = 0.147). The response type also interacted with congruency in right anterior insula ( F (1,21) = 4.666, p = 0.042, 𝜂2 𝑝 = 0 . 182 ). Trials with the response “same ”induced higher response than the trials with the response “different ”in the incongruent condition ( p < 0.001), but did not differ from each other in the congruent condition ( p = 0.370). There was no response by congruency interaction in left anterior insula ( F (1,21) = 3.473, p = 0.076), nor mPFG ( F (1,21) = 0.504, p = 0.485). As for the other two regions identified by Eriksen Flanker task, namely the left inferior parietal lobule and left precentral gyrus, there was no main effect of alignment, congruency, or response, nor interaction between any two factors or three factors (p > 0.05 uncorrected). Importantly, there were three-way interactions in right anterior insula ( F (1,21) = 7.239, p = 0.014, 𝜂2 𝑝 = 0 . 256 ) and mPFG ( F (1,21) = 5.042, p = 0.036, 𝜂2 𝑝 = 0 . 194 ), though both failed surviving the Bonferroni correction for 10 ROIs. Simple effect analysis showed that in these areas, there was a significant interaction between congruency and response type in aligned faces (right anterior insula, F (1,21) = 9.058, p = 0.007, 𝜂2 𝑝 = 0 . 301 , mPFG ( F (1,21) = 4.121, p = 0.055 , 𝜂2 𝑝 = 0 . 164 ), but not in misaligned faces (right anterior insula, F (1,21) = 0.184, p = 0.672, mPFG, F (1,21) = 0.909, p = 0.351). The interaction in aligned faces was characterized by higher BOLD response in incongruent condition over the congruent condition when the top halves were “the same ” (right anterior insula ( F (1,21) = 25.93, p < 0.001 , 𝜂2 𝑝 = 0 . 552 ), and mPFG ( F (1,21) = 18.487, p < 0.001, 𝜂2 𝑝 = 0 . 468) , but no congruency effect when the top halves were “different ”(right anterior insula, F (1,21) = 0.249, p = 0.623, mPFG, F (1,21) = 0.557, p = 0.464). 3.4. Behavioral-neural correspondence in ROIs The results of multidimensional scaling analysis and hierarchical clustering on BOLD signal in the ten ROIs are available in Fig. 5 . Hierarchical clustering revealed three consistent clusters among the 8 conditions in areas including the bilateral insula and the mPFG. In all three 7 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Fig. 5.. The clustering revealed by dendrograms among the experimental conditions in each ROI. The right insula, left insula and medial prefrontal gyrus demonstrate similar proximity structure among the 8 experimental conditions as in the behavioral response. Fig. 6.. Distribution of the 8 conditions along the two principal dimensions by the multidimensional scaling analysis. blue dots: conditions in behavioral data, red dots: conditions in neural data. areas, the first cluster included only one condition: (AIS), which was distant from all the other conditions. The second cluster consisted of the ACD, MCD and MID conditions, where the correct response was all “different ”. The third condition contained the ACS, AID, MCS and MIS conditions, where the correct response was mostly “the same ” except the AID condition. The patterns of clustering in the other areas did not convey explicit information either about response type or attentional control. Multidimensional scaling analysis on dissimilarity matrix extracted two major dimensions among neural response of these 8 conditions in bilateral insula and the mPFG. As can be seen in Fig. 6 , the first dimension generated by the multidimensional scaling analysis was largely characterized by the distance of the AIS condition away from all other 7 conditions. The second dimension reflected mainly the type of response, with most of the conditions with the response “different ”at one end and most conditions with the response “same ”at the other end. To explore brain regions where the neural activities fit satisfactorily to the behavioral results, we estimated the behavioral and neural dissimilarity matrices of the 8 conditions, and matched them by the Spearman’s rank correlation. Permutation simulation analysis ( n = 10,000) on the correlation, as plotted in Fig. 6 , showed significant neural-behavioral correspondence in the right anterior insula ( p < 0.001), left anterior insula ( p < 0.001), both survived the Bonferroni multiple correction for N = 10 ROIs. The neural pattern in mPFG also matched that of the behavioral results ( p = 0.006), with a significance trending to q < 0.05 by Bonferroni correction. In contrast, the neural-behavioral conformation was not significant in any of the 5 face processing regions (for right OFA, p = 0.113, right STS, p = 0.319, left FFA, p = 0.391, left STS, p = 0.329, right FFA, p = 0.856). These results suggested the neural response in the bilateral insula areas showed better fitting to the pattern in the behavioral data than areas in the face processing network. It is also worth noting that similar results were obtained when the reference behavioral dissimilarity was estimated based on response accuracy (see supplementary information). To assess the reliability of clustering across participants, we used a n-fold cross-validation technique ( Chen et al., 2015 ; Qin et al., 2014 ) based on the dissimilarity indices as recommended in previous stud8 C. Chen, Y. Lou, H. Li et al. NeuroImage 246 (2022) 118756 Qin, S., Young, C.B., Duan, X., Chen, T., Supekar, K., Menon, V., 2014. Amygdala subregional structure and intrinsic functional connectivity predicts individual differences in anxiety during early childhood. Biol. Psychiatry 75 (11), 892–900. doi: 10.1016/j.biopsych.2013.10.006 . Ramon, M., Busigny, T., Rossion, B., 2010. Impaired holistic processing of unfamiliar individual faces in acquired prosopagnosia. Neuropsychologia 48 (4), 933–944. doi: 10.1016/j.neuropsychologia.2009.11.014 . Renzi, C., Schiavi, S., Carbon, C., Vecchi, T., Silvanto, J., Cattaneo, Z., 2013. Processing of featural and configural aspects of faces is lateralized in dorsolateral prefrontal cortex: a TMS study. Neuroimage 74, 45–51. doi: 10.1016/j.neuroimage.2013.02.015 . Rezlescu, C., Susilo, T., Wilmer, J.B., Caramazza, A., 2017. The inversion, part-whole, and composite effects reflect distinct perceptual mechanisms with varied relationships to face recognition. J. Exp. Psychol. Hum. Percept. Perform. 43 (12), 1961–1973. doi: 10.1037/xhp0000400 . Richler, J.J., Cheung, O.S., Gauthier, I., 2011. Holistic processing predicts face recognition. Psychol. Sci. 22 (4), 464–471. doi: 10.1177/0956797611401753 . Richler, J.J., Cheung, O.S., Wong, A.C., Gauthier, I., 2009. Does response interference contribute to face composite effects? Psychon. Bull. Rev. 16 (2), 258–263. doi: 10.3758/PBR.16.2.258 . Richler, J.J., Gauthier, I., 2013. When intuition fails to align with data: a reply to Rossion (2013). Vis. Cogn. 21 (2). doi: 10.1080/13506285.2013.796035 . Richler, J.J., Gauthier, I., Wenger, M.J., Palmeri, T.J., 2008. Holistic processing of faces: perceptual and decisional components. J. Exp. Psychol. Learn. Mem. Cogn. 34 (2), 328–342. doi: 10.1037/0278-7393.34.2.328 . Richler, J.J., Palmeri, T.J., Gauthier, I., 2012. Meanings, mechanisms, and measures of holistic processing. Front. Psychol. 3 (DEC). doi: 10.3389/fpsyg.2012.00553 . Ross, D.A., Tamber-Rosenau, B.J., Palmeri, T.J., Zhang, J., Xu, Y., Gauthier, I., 2018. Highresolution functional magnetic resonance imaging reveals configural processing of cars in right anterior fusiform face area of car experts. J. Cogn. Neurosci. 30 (7), 973–984. doi: 10.1162/jocn_a_01256 . Rossion, B., 2013. The composite face illusion: a whole window into our understanding of holistic face perception. Vis. Cogn. 21 (2), 139–253. doi: 10.1080/13506285.2013.772929 . Rotshtein, P., Geng, J.J., Driver, J., Dolan, R.J., 2007. Role of features and second-order spatial relations in face discrimination, face recognition, and individual face skills: behavioral and functional magnetic resonance imaging data. J. Cogn. Neurosci. 19 (9), 1435–1452. doi: 10.1162/jocn.2007.19.9.1435 . Schiltz, C., Dricot, L., Goebel, R., Rossion, B., 2010. Holistic perception of individual faces in the right middle fusiform gyrus as evidenced by the composite face illusion. J. Vis. 10 (2), 1–16. doi: 10.1167/10.2.25 . Schiltz, C., Rossion, B., 2006. Faces are represented holistically in the human occipito-temporal cortex. Neuroimage 32 (3), 1385–1394. doi: 10.1016/j.neuroimage.2006.05.037 . Soria Bauser, D.A., Suchan, B., Daum, I., 2011. Differences between perception of human faces and body shapes: evidence from the composite illusion. Vis. Res. 51 (1), 195– 202. doi: 10.1016/j.visres.2010.11.007 . Stanislaw, H., Todorov, N., 1999. Calculation of signal detection theory measures. Behav. Res. Methods Instrum. Comput. 31 (1), 137–149. doi: 10.3758/bf03207704 . Susilo, T., McKone, E., Dennett, H., Darke, H., Palermo, R., Hall, A., Rhodes, G., 2010. Face recognition impairments despite normal holistic processing and face space coding: evidence from a case of developmental prosopagnosia. Cogn. Neuropsychol. 27 (8), 636–664. doi: 10.1080/02643294.2011.613372 . Susilo, T., Rezlescu, C., Duchaine, B., 2013. The composite effect for inverted faces is reliable at large sample sizes and requires the basic face configuration. J. Vis. 13 (13), 14. doi: 10.1167/13.13.14 , -14 . Talati, A., Hirsch, J., 2005. Functional specialization within the medial frontal gyrus for perceptual go/no-go decisions based on "what," "when," and "where" related information: an fMRI study. J. Cogn. Neurosci. 17 (7), 981–993. doi: 10.1162/0898929054475226 . Tanaka, J.W., Farah, M.J., 1993. Parts and wholes in face recognition. Q. J. Exp. Psychol. Sect. A 46 (2), 225–245. doi: 10.1080/14640749308401045 . Tanaka, J.W., Simonyi, D., 2016. The "parts and wholes" of face recognition: a review of the literature. Q. J. Exp. Psychol. 69 (10), 1876–1889. doi: 10.1080/17470218.2016.1146780 , (2006) . Todorov, A., Loehr, V., Oosterhof, N.N., 2010. The obligatory nature of holistic processing of faces in social judgments. Perception 39 (4), 514–532. doi: 10.1068/p6501 . Tso, R.V. , Chen, H. , Yeung, Y.A. , Au, T. , Hsiao, J. , 2017. Right hemisphere lateralization and holistic processing do not always go together: an ERP investigation of a training study. Cogn. Sci. 3386–3391 . Tzagarakis, C., Jerde, T.A., Lewis, S.M., U ğurbil, K., Georgopoulos, A.P., 2009. Cerebral cortical mechanisms of copying geometrical shapes: a multidimensional scaling analysis of fMRI patterns of activation. Exp. Brain Res. 194 (3), 369–380. doi: 10.1007/s00221-009-1709-5 . Uddin, L.Q., Nomi, J.S., Hébert-Seropian, B., Ghaziri, J., Boucher, O., 2017. Structure and function of the human insula. J. Clin. Neurophysiol. 34 (4), 300–306. doi: 10.1097/wnp.0000000000000377 . Ullsperger, M., Harsay, H.A., Wessel, J.R., Ridderinkhof, K.R., 2010. Conscious perception of errors and its relation to the anterior insula. Brain Struct. Funct. 214 (5), 629–643. doi: 10.1007/s00429-010-0261-1 . Utevsky, A.V., Smith, D.V., Huettel, S.A., 2014. Precuneus is a functional core of the default-mode network. J. Neurosci. 34 (3), 932–940. doi: 10.1523/jneurosci.4227-13.2014 . Varoquaux, G., 2018. Cross-validation failure: small sample sizes lead to large error bars. Neuroimage 180, 68–77. doi: 10.1016/j.neuroimage.2017.06.061 . Ventura, P., Leite, I., Fernandes, T., 2018. The development of holistic face processing: an evaluation with the complete design of the composite task. Acta Psychol. (Amst) 191, 32–41. doi: 10.1016/j.actpsy.2018.08.015 . Von Der Heide, R.J., Wenger, M.J., Bittner, J.L., Fitousi, D., 2018. Converging operations and the role of perceptual and decisional influences on the perception of faces: neural and behavioral evidence. Brain Cogn. 122 (October 2017), 59–75. doi: 10.1016/j.bandc.2018.01.007 . Walther, A., Nili, H., Ejaz, N., Alink, A., Kriegeskorte, N., Diedrichsen, J., 2016. Reliability of dissimilarity measures for multi-voxel pattern analysis. Neuroimage 137, 188–200. doi: 10.1016/J.NEUROIMAGE.2015.12.012 . Wang, R., Li, J., Fang, H., Tian, M., Liu, J., 2012. Individual differences in holistic processing predict face recognition ability. Psychol. Sci. 23 (2), 169–177. doi: 10.1177/0956797611420575 . Wang, X., Zhu, Q., Song, Y., Liu, J., 2017. Developmental reorganization of the core and extended face networks revealed by global functional connectivity. Cereb. Cortex 28 (10), 3521–3530. doi: 10.1093/cercor/bhx217 . Wang, Y., Guo, S., Sun, Q., Jin, S.F., Zhang, X., Xiao, M., Li, K.Z., 2019. Anthropometric labial analysis of Han Chinese young adults. Skin Res. Technol. 25 (4), 499–503. doi: 10.1111/srt.12678 . Wong, A.C., Bukach, C.M., Hsiao, J., Greenspon, E., Ahern, E., Duan, Y., Lui, K.F., 2012. Holistic processing as a hallmark of perceptual expertise for nonface categories including Chinese characters. J. Vis. 12 (13), 7. doi: 10.1167/12.13.7 . Wong, A.C., Palmeri, T.J., Rogers, B.P., Gore, J.C., Gauthier, I., 2009. Beyond shape: how you learn about objects affects how they are represented in visual cortex. PLoS One 4 (12). doi: 10.1371/journal.pone.0008405 . Wu, T., Schulz, K.P., Fan, J., 2021. Activation of the cognitive control network associated with information uncertainty. Neuroimage 230, 117703. doi: 10.1016/j.neuroimage.2020.117703 . Wu, T., Spagna, A., Chen, C., Schulz, K.P., Hof, P.R., Fan, J., 2020. Supramodal mechanisms of the cognitive control network in uncertainty processing. Cereb. Cortex 30 (12), 6336–6349. doi: 10.1093/cercor/bhaa189 . Wu, T., Wang, X., Wu, Q., Spagna, A., Yang, J., Yuan, C., Fan, J., 2019. Anterior insular cortex is a bottleneck of cognitive control. Neuroimage 195, 490–504. doi: 10.1016/j.neuroimage.2019.02.042 . Xie, W., Zhang, W., 2016. The influence of emotion on face processing. Cogn. Emot 30 (2), 245–257. doi: 10.1080/02699931.2014.994477 . Yang, L.Z., Zhang, W., Shi, B., Yang, Z., Wei, Z., Gu, F., Rao, H., 2014. Electrical stimulation over bilateral occipito-temporal regions reduces N170 in the right hemisphere and the composite face effect. PLoS One 9 (12). doi: 10.1371/journal.pone.0115772 . Yin, R.K., 1969. Looking at upside-down faces. J. Exp. Psychol. 81 (1), 141–145. doi: 10.1037/H0027474 . Young, A.W., Hellawell, D., Hay, D.C., 1987. Configurational information in face perception. Perception 16 (6), 747–759. doi: 10.1068/p160747 . Zhang, H., Tian, J., Liu, J., Li, J., Lee, K., 2009. Intrinsically organized network for face perception during the resting state. Neurosci. Lett. 454 (1), 1–5. doi: 10.1016/j.neulet.2009.02.054 . Zhang, J., Liu, J., Xu, Y., 2015. Neural decoding reveals impaired face configural processing in the right fusiform face area of individuals with developmental prosopagnosia. J. Neurosci. 35 (4), 1539–1548. doi: 10.1523/jneurosci.2646-14.2015 . Zhao, M., Bülthoff, H.H., Bülthoff, I., 2015. Beyond faces and expertise: facelike holistic processing of nonface objects in the absence of expertise. Psychol. Sci. 27 (2), 213– 222. doi: 10.1177/0956797615617779 . Zhao, M., Bülthoff, H.H., Bülthoff, I., 2016. A shape-based account for holistic face processing. J. Exp. Psychol. Learn. Mem. Cogn. 42 (4), 584–597. doi: 10.1037/xlm0000185 . Zhao, M., Hayward, W.G., Bülthoff, I., 2014. Holistic processing, contact, and the other-race effect in face recognition. Vis. Res. 105, 61–69. doi: 10.1016/j.visres.2014.09.006 . 15