The anterior fusiform gyrus: The ghost in the cortical face machine

Face-selective regions in the human ventral occipito-temporal cortex (VOTC) have been defined for decades mainly with functional magnetic resonance imaging. This face-selective VOTC network is traditionally divided in a posterior ‘core ’ system thought to subtend face perception, and regions of the anterior temporal lobe as a semantic memory component of an extended general system. In between these two putative systems lies the anterior fusiform gyrus and surrounding sulci, affected by magnetic susceptibility artifacts. Here we suggest that this methodological gap overlaps with and contributes to a conceptual gap between (visual) perception and semantic memory for faces. Filling this gap with intracerebral recordings and direct electrical stimulation reveals robust face-selectivity in the anterior fusiform gyrus and a crucial role of this region, especially in the right hemisphere, in identity recognition for both familiar and unfamiliar faces. Based on these observations, we propose an integrated theoretical framework for human face (identity) recognition according to which face-selective regions in the anterior fusiform gyrus join the dots between posterior and anterior cortical face memories.


The cortical network for human face recognition
A fundamental function of central nervous systems is to adaptively recognize stimuli in their physical and social environment.Whereas in Psychology the term "recognition" often implies a judgment of previous occurrence (specifically "the ability to identify information as having been encountered before", APA Dictionary of Psychology; see also Mandler, 1980), recognition can be more broadly defined in biology as the production of a selective (i.e., discriminant) response to a given signal, a response that is reliable (i.e., reproducible) across a wide variety of instances of that signal (Edelman, 2004).In the human species, recognition of signals from the face of conspecifics is essential for social interactions.Neurotypical human adults recognize people's gender, emotional expression, age, ethnical origin, identity etc. from their face with remarkable efficiency, speed and automaticity (Calder et al., 2011;Bruce and Young, 2023).
The neural basis of this human face recognition function, at large, has been investigated extensively with functional neuroimaging for 30 years: first with positron emission tomography (PET; Sergent et al., 1992) thenand mainlywith functional magnetic resonance imaging (fMRI; since Puce et al., 1995).Collectively, these studies have disclosed clusters of voxels of a few cubic millimeters with larger neural activation to pictures of faces than nonface visual stimuli in the human brain (e.g., Weiner and Grill-Spector, 2010;Rossion et al., 2012;Zhen et al., 2015; see Behrmann et al., 2016;Grill-Spector et al., 2017 for reviews).Despite a substantial amount of variability in the paradigms and stimuli used across fMRI studies (Duncan et al., 2009;Berman et al., 2010), these face-selective clusters have been consistently reported in both the Ventral Occipito-Temporal Cortex (VOTC) and the Superior Temporal Sulcus (STS) (Fig. 1A, B).
The fMRI-defined face-selective clusters of the human occipitotemporal cortex have been labelled according to the anatomical region where they are usually observed.For instance, the well-known "Fusiform Face Area" ("FFA", labelled by Kanwisher et al., 1997) is a face-selective cluster identified in the lateral portion of middle section along the anterior-posterior axis of the fusiform gyrus (midFG), while the "Occipital Face Area" ("OFA", labelled by Gauthier et al., 2000) is typically identified in the lateral portion of the inferior occipital gyrus or IOG (Fig. 1A, B).Following this logic, up to 6 face-selective clusters, namely 4 in the VOTC (OFA, pFFA [posterior FFA, also called pFus-faces or FFA-1], mFFA [middle FFA or mFus-faces, or FFA-2], ATL-FA [Anterior Temporal Lobe Face Area]) and 2 or 3 in the STS (Fig. 1B), have been defined in the most recent neurofunctional model of human face recognition (Duchaine and Yovel, 2015; for earlier models: Haxby et al., 2000;Calder and Young, 2005;Haxby and Gobbini, 2011).These face-selective clusters of voxels are thought to contain populations of (millions of) neurons (Logothetis, 2008) which, by definition, must play a key role in the recognition of a visual stimulus as a face.Beyond this generic face recognition function, many fMRI studies have tested the sensitivity of these face-selective clustersin particular the FFA -to physical stimulus manipulations (e.g., position, size, head orientation of faces and various image statistics; e.g., Tong et al., 2000;Rice et al., 2014;Finzi et al., 2021), to attention (e.g., Peelen et al., 2009) as well as conscious perception (e.g., Tong et al., 1998;Fang and He, 2005), and have investigated their putative role in finer-grained facial recognition functions (e.g., face familiarity and identity, facial expression, eye gaze direction, etc.; for reviews see Haxby and Gobbini, 2011;Rossion, 2014;Duchaine and Yovel, 2015;Grill-Spector et al., 2017).Another important line of research focuses on the origin and developmental trajectory of these fMRI face-selective clusters (e.g., Golarai et al., 2007;Behrmann et al., 2016;Golarai et al., 2017;Kosakowski et al., 2022; see Scott and Arcaro, 2023 for review).
Overall, the ultimate objectives of this research program are to (1) define each component of the human cortical face network, (2) determine its anatomical features and intrinsic/extrinsic anatomicofunctional connections and (3) understand the nature of its local representations and processes (Grill-Spector et al., 2017).For instance, fMRI studies have associated face-selective regions of the human STS with dynamic aspects of face recognition (e.g., facial expression, eye gaze and head orientation) while those in the VOTC would rather be predominantly involved in more stable aspects of face recognition (e.g., identity, gender, etc.) (Haxby et al., 2000;Duchaine and Yovel, 2015).In both the 'dorsal' (STS) and ventral (VOTC) pathways, the prevalent view is that of a progressive, hierarchical, evolution in the degree of invariance (e.g.size, position, head orientation,…) and complexity of facial representations from posterior to anterior face-selective regions (Duchaine and Yovel, 2015;Grill-Spector et al., 2018;Tsantani et al., 2021).Inspired by Hubel and Wiesel (1977)'s description of the macaque visual system as a "cortical machinery", this apparent structured set of local domain-specific processing units interconnected and interacting in a definite order to collectively achieve a specific complex function, namely face recognition, has been often referred to as a "face machinery" (Kanwisher, 2010;Tsao and Livingstone, 2008;Yovel and Freiwald, 2013).

The ghost in the machine
Human neuroimaging studies using face stimuli have most extensively investigated regions of the VOTC, i.e., the ventral cortical face network, in particular the (right) FFA in the midFG arguably the most consistent category-selective neural cluster in the human brain (Kanwisher, 2017).One reason for the emphasis on the VOTC is its key role in face identity recognition (FIR), as shown by lesion studies of a category-selective impairment in FIR, prosopagnosia (Meadows, 1974;Bouvier, Engel, 2006;Cohen et al., 2019;Rossion, 2022a;Rossion, 2022b).Yet, strikingly, as described in neuroimaging studies, reviews and functional models, the ventral cortical face network appears to stop abruptly in the postero-anterior VOTC axis at the level of the midFG, as if the mFus-faces/FFA-2 region was the end point of the network (Fig. 1A).
Admittedly, in the last decade or so, face-selective clusters in more anterior regions of the VOTC, i.e., in the ventral anterior temporal lobe (ATL), have also been reported in a handful of studies (ATL-faces or ATL-FA, Rajimehr et al., 2009;Nasr and Tootell, 2012;Rossion et al., 2012;Avidan et al. 2014;Collins and Olson, 2014;Collins et al., 2016, Fig. 1B).These clusters are relatively small, scattered and found in only a subset of individual brains (e.g., 15/36 participants in Rossion et al., 2012; 3/10 hemispheres in Pyles et al., 2013).They are located very anteriorly (i.e., close to the temporal pole) and, contrary to clusters of the posterior VOTC cortical face network (Weiner and Grill-Spector, 2010;Grill-Spector et al., 2017), these ventral ATL face clusters do not appear to have a consistent anatomical location across studies.
In summary, an overview of fMRI-based face-selective activity in the human brain points to the FFA/mFus-faces as an end point, or at least to a substantial spatial gap in between the FFA/mFus-faces and the ATL face areas close to the temporal pole (Fig. 1; Fig. 2).It is as if the faceselective computations in the "cortical face machine" were completed at the level of the FFA/mFus-faces, and the output of these computations had to be sent to a spatially distant extended system of anteriorlylocated brain regions in the temporal lobe (Haxby et al., 2000;Calder and Young, 2005;Gobbini and Haxby, 2007;Duchaine and Yovel, 2015;Kovacs, 2020) (Fig. 1B, Fig. 3A); see also structural and functional connectivity studies (Gshwind et al., 2012;Pyles et al., 2013;Gomez et al., 2015;Wang et al., 2020; Fig. 1C).

A methodological gap
Anatomically, the gap in the cortical face system is located just anteriorly to the midFG and the mid-fusiform sulcus, at the level of the anterior section of the fusiform gyrus (antFG) and adjacent sulci, the anterior collateral sulcus (antCOS) and anterior occipito-temporal sulcus (antOTS) (Fig. 2A).For sake of simplicity, these latter 3 regions considered altogether will be named antFG+ hereafter.
Why would this antFG+ region not be selective to face stimuli, given that face-selective responses are measured all along the more posterior section of the FG?Why is there an apparent gap at this level in the ventral cortical face network?Does it correspond to a real functional gap, i.e., a cortical region or set of regions not involved in face recognition, or is this gap artificial?
In fMRI, the anterior temporal lobe, and in particular the antFG+ region, is affected by a large susceptibility artifact arising from the auditory canals (Fig. 2).In the presence of additive instrumental noise, the signal-to-noise ratio (SNR) will be very inhomogeneous, rendering this region notoriously difficult to measure (e.g., Devlin et al., 2000).Thus, most fMRI reports, using conventional gradient-echo Echo-Planar Imaging (EPI), are largely blind to cortical responses in the portion of VOTC near the auditory canal (Ojemann et al., 1997;Deng et al., 2009;Wandell, 2011).Since the anterior portion of the identified FFA/mFus-faces abuts this artifact-ridden region, and there is a significant decrease of SNR between the midFG and the antFG (Zhang et al., 2016), it is entirely possible that genuine category-selective neural responses in the antFG+ region cannot be detected because of the artifact (Fig. 2).
Despite this artifact being well-known and having been explicitly acknowledged as a severe limitation to map the neural basis of other domains such as semantics and visual word recognition with fMRI (e.g., Devlin et al., 2000;Visser et al., 2010;Wandell, 2011;Halai et al., 2014;Persichetti et al., 2021), it is largely ignored in human face recognition research.Moreover, when it is acknowledged, it is not adequately dealt with.For instance, Axelrod and Yovel (2013) proposed to use coronal slices as a solution to eliminate or reduce the magnetic susceptibility artifact in the ventral ATL, claiming that this procedure would be able to identify cortical face-selective regions with adequate SNR in this region.However, to our knowledge, these authors' proposal has not been considered by other researchers in the field of human face recognition, and has not been validated.Indeed, and most importantly, the antFG+ , where the main magnetic susceptibility artifact is found, was not sampled in the study of Axelrod and Yovel (2013), perhaps under the assumption that this region indeed does not contain any face-selective region of interest.
In the past decade, distortion-corrected fMRI sequences have been developed to improve SNR in regions affected by magnetic susceptibility artifacts, in particular the ventral ATL (Embleton et al., 2010;Visser et al., 2010).These fMRI sequences, e.g. using dual-echo EPI imaging or spin-echo EPI imaging (Halai et al., 2014), have been used in particular to probe the neural basis of multimodal semantics in the ventral ATL (e. g., Visser et al., 2010;Binney et al., 2016), sometimes with pictures of faces or face localizers (Rice et al., 2018).Yet, even with these distortion-corrected fMRI sequences, SNR remains relatively low specifically in the antFG+ region (Collins et al., 2016; see e.g., Fig. 1 in Hoffman et al., 2015;Fig. 2 in Rice et al., 2018; see discussion in Rossion et al., 2018) limiting its deeper exploration and understanding with this Fig. 2. The methodological gap involving the anterior section of the fusiform gyrus (antFG).A. Anatomical location of the antFG displayed on a cortical reconstruction of the Colin27 brain (in red).The antFG is located anteriorly to the midFG (divided in latFG and medFG by the mid-fusiform sulcus, MFS; see Weiner, 2019) and to a plane defined by the posterior tip of the hippocampus (dotted black line).B. In fMRI, the ventral ATL is affected by a severe signal drop-out primarily involving the antFG, as shown by the area in black indicating very low MRI signal intensity (from Shum et al., 2013, with permission).C. Examples of MRI signal intensity and tSNR together with the location of face-selective activations displayed on ventral views of inflated cortical surfaces.The vATL signal drop-out is indicated by a white arrow (gray or black zone in Rajimehr et al., 2009 andin Volfart et al., 2022;red zone in Lafer-sousa et al., 2016, with permission).Face-selective activations (green outlines in Rajimehr et al., 2009, purple in Lafer-sousa et al., 2016, white in Volfart et al., 2022) are found all along the VOTC except in the region of the signal drop-out.Abbreviations: ant, anterior; COS, collateral sulcus; FG, fusiform sulcus; lat, lateral; med, medial; MFS, mid-fusiform sulcus; OTS, occipito-temporal sulcus; TP, temporal pole.

(caption on next page)
B. Rossion et al. technique.Moreover, in general, the fMRI community has not yet set standards for communicating the SNR of the measurement, and there are no known methods that equate SNR along the entire VOTC (Wandell, 2011).Therefore, with a uniform threshold-based approach used across the VOTC, the antFG+ will always lag behind both face-selective posterior VOTC regions (i.e., OFA, FFA) and regions closer to the temporal pole in terms of their putative importance for human face recognition.

Mind the gap: perception vs. memory
What if the disproportionate focus on the FFA in the community of fMRI researcherss was not due to this midFG region generating the most consistent face-selective activity but simply because (more) consistent anterior face-selective responses were masked by the methodological artifact?More generally, what are the implications of this artifact at the level of the antFG+ for our conceptualization of the neural basis of human face recognition?Is the gap in the human ventral cortical face network only about methodological issues?
As noted above, it seems that the methodological gap is largely ignored in the scientific community working on the neural basis of human face recognition.Moreover, even when they are acknowledged, anterior temporal lobe face-selective regions are considered as being outside or beyond the core face network, as if there was a fundamental gap in function between posterior and anterior VOTC regions.This view is exemplified in the extensive review on the functional neuroanatomy of human face perception by Grill-Spector and colleagues (2017) in which the authors state from the outset that "Additional face-selective regions have been identified in the anterior temporal lobe … but these regions are not considered part of the core face network … as they are not just driven by visual stimulation and are more elusive to identify because of lower signals and susceptibility artifacts in fMRI" (Grill-Spector et al., 2017, p.170).What is meant by "not driven just by visual stimulation" is unspecified, but the authors' reasoning is in line with the above-described hypothetical division of labor between a core system for face perception in posterior regions (STS, OFA and FFA complex) and a more anterior extended system including anterior and medial temporal lobe structures thought to be involved in generic semantic, episodic and affective memory functions (Haxby et al., 2000;Calder and Young, 2005;Haxby and Gobbini, 2011;Duchaine and Yovel, 2015) (Fig. 3).
Based on these considerations, the point that we want to make here is that the gap in the VOTC map of human face recognition does not only anatomically overlap with a hypothetical border between core regions (posterior VOTC) and extended regions (ATL) of the face network, but it also creates, and fuels, a rift between the two general cognitive functions thought be involved: perception and memory.As put forward by Wandell (2011, p.71-72) in a different context "rather than being acknowledged as instrumental limitations, weak or absent (fMRI) signals are interpreted as if they are properties of the brain and incorporated as part of the theory".
In human face recognition research, the conceptual gap between 'perception' and 'memory' is substantial, reflecting the general view that the face recognition function would be divided in two main stages: first a series of (hierarchical) information processes leading to the construction of invariant visual representations ("structural encoding codes") that are, at a second stage, associated with representations of (familiar) faces stored in memory ('facial recognition units') and multimodal semantics (Fig. 3).This major division of labor is an essential feature of standard cognitive models of human face recognition (Bruce and Young, 1986;Young and Bruce, 2011; see also Fig. 1 in Young, 2021), and implemented explicitly in every neuro-cognitive model of this function (Haxby et al., 2000;Calder and Young, 2005;Pitcher et al., 2011;Haxby and Gobbini, 2011;Duchaine and Yovel, 2015; see also Freiwald, 2020;Hesse and Tsao, 2020).More generally, visual object recognition is generally conceptualized in cognitive/computational science as being based on the same hierarchical distinction between perceptual and (semantic) memory stages (Marr, 1982;Humphreys & Riddoch, 1987;Firestone and Scholl, 2016;Riesenhuber and Poggio, 1999).This distinction is based on an conceptual analysis of the (visual) recognition process (Marr, 1982;DiCarlo et al., 2012), and is rooted in the classical neuropsychological distinction of Lissauer (1890) between apperceptive and associative forms of visual agnosia, subsequently adopted in human face recognition: while (brain-damaged) patients with apperceptive prosopagnosia are thought to present with deficits in building correct visual representations of faces, patient with associative prosopagnosia would be unable to associate their correctly built percept with memory representations of these faces (i.e., a face percept "stripped of its meaning"; Hécaen, 1981;De Renzi, 1986;De Renzi et al., 1991;Davies-Thompson et al., 2014;Barton and Corrow, 2016; see Fig. 3).While the localization of lesions causing the putative associative form of prosopagnosia remains unclear, it has even been hypothetically schematized at the junction between regions thought to deal with perceptual processes (OFA & FFA) and those involved in semantic memory (temporal pole), roughly at the location of the antFG+ (Davies-Thompson et al., 2014; Fig. 3B).

Joining the dots of the cortical face network
Is there really a gap in face-selective activity in the human VOTC at the level of the antFG+ and does it truly correspond to a neat distinction between perceptual and memory stages of human face recognition ?Since the fusiform gyrus is a hominoid structure (Bryant and Preuss, 2018;Roumazeilles et al., 2020;Parker et al., 2023) and macaque monkeys do not show face-selective activity in their posterior VOTC (e. g., Tsao et al., 2008;Laurent et al., 2023;Fig. 4), neuroimaging or electrophysiological recordings of the cortical face network in this non-human primate species cannot address this issue (Rossion and Taubert, 2019;Rossion, 2022c).Rather, one must turn to other methodologies of investigation of the human brain.
Several early neuroimaging studies with PET, for which there is no issue of signal drop-out in the ventral ATL, reported antFG activation for pictures of faces.Kuskowski and Pardo (1999) reported activation in the bilateral antFG for clear vs. scrambled (meaningless) images of faces, with the region clearly demarcated from the (also activated) right midFG.Two subsequent studies identified the right antFG in a contrast between familiar and unfamiliar faces (Rossion et al., 2001;Wiser et al., 2000).However, none of these studies provided evidence of category-selectivity in this region (i.e., larger or differential response to faces vs. other significant visual forms).
Today, the only alternative approach to fMRI to achieve a comprehensive map of face-selectivity in humans is provided by electrophysiological recordings in awake patients implanted with intracranial electrodes along the VOTC as part of their presurgical evaluation for drug-resistant focal epilepsy.These relatively rare-compared with neuroimaging-intracranial EEG (iEEG) recordings allow millisecond Fig. 3.A gap between perception and memory in neuro-cognitive models of human face recognition.A. Bottom: The influential model of Haxby et al. (2000) along with the schematic locations of each region as shown in the recent review of Kovács (2020) (reproduced with permission).Note the dotted lines separating the core system thought to be involved in perception or visual analysis (yellow) and the extended system, as well as the spatial gap between the ventral core regions (OFA and FFA) and the ATL region involved in storing semantic identity-specific information.This gap corresponds to the location of the antFG+ (star).B. A schematic cognitive decomposition of the human face recognition function, along with proposed locations of each region subtending these subfunctions and the putative consequences on face recognition when these regions are damaged (Davies-Thompson et al., 2014, with permission).In this hierarchical model, the face percept generated by perceptual stages (blue: OFA, FFA) needs to successfully match a stored facial memory in the ventral ATL (red: anterior inferior temporal or aIT) to activate semantic identity-specific information.resolution measurements of electrical fields directly associated with local neural activity, but mostly importantly for the present issue, with high SNR and stable noise levels across VOTC regions (see Jacques et al., 2022).
Early human iEEG studies measuring event-related potentials (ERP) for faces and objects have shown face-selective activity all along the VOTC, without any gap (Allison et al., 1999;Puce et al., 1999).However, there was no systematic quantification of face-selectivity across anatomical regions in these early studies.Moreover, these studies were performed with subdural electrodes applied onto the gyral cortical surface (electrocorticography, ECoG), a surgical technique for intracranial electrode placement that is limited in its exploration of cortical sulci (i.e., two-third of the cortex in humans; Zilles et al., 1988).
To our knowledge, a systematic measurement of face-selectivity in the human VOTC has been performed only in the past decade with a combination of intracerebral "depth" electrodes inserted inside the brain (stereo-electroencephalography, SEEG) and a frequency-tagging approach (fast periodic visual stimulation, FPVS; Rossion et al., 2018, see Fig. 5A; see Jonas et al., 2016;Hagen et al., 2020;Jacques et al., 2022).According to these SEEG-FPVS studies, face-selective activity is found all along the VOTC, from the IOG to the temporal pole, without any discontinuity (Fig. 5B).Importantly, there is no decrease of face-selectivity at the level of the antFG+ (see Jonas et al., 2016 for examples of face-selective activities in the antFG+ of an individual brain).
Despite this continuity, there are focal points of larger amplitudes, first and foremost in the lateral section of the midFG with a right predominanceconfirming with direct measures of brain activity the predominant role of this 'FFA' regionbut also in the IOG and, importantly for the present review, the antFG+ (Fig. 5C).Moreover, face-selectivity decreases from the antFG+ to the temporal pole, so that, contrary to what is reported in fMRI studies, face-selectivity is in fact higher in the antFG+ at the level of the usual location of the maximum signal drop-out than in the temporal pole where the ATL-FA cluster(s) is (are) sometimes found in fMRI (Fig. 5C).Importantly, and despite an overall lower face-selective signal particularly in the anterior VOTC, these observations are also valid when measuring high-frequency activity (>30 Hz also known as 'gamma' or 'high-frequency broadband') with the FPVS paradigm illustrated on Fig. 5 (see Jacques et al., 2022).
In summary, the findings of a continuum of intracerebral faceselectivity all along the VOTC up to the temporal pole goes against this putative gap between "perception" and "memory" functions in human face recognition.

Is the antFG critical for face identity recognition?
Having established the antFG+ as a reliable face-selective region thanks to human intracerebral recordings, an outstanding issue is whether this region is critical for face recognition and if so, what is the nature of its contribution?
To our knowledge, there is no case study reported in the literature with a lesion restricted to the antFG+ , a relatively small cortical region.Interestingly, low performers at recognizing face identity without neurological history (i.e., developmental prosopagnosia/prosopdysgnosia; Rossion, 2018;Sørensen and Overgaard, 2018) may show a volume reduction of their bilateral antFG correlated with their behavioral decrement in famous face identity recognition (Behrmann et al., 2007).A similar observation has been reported in patients with fronto-temporal lobar degeneration (temporal and frontal variants), with a specific correlation with the right antFG (Omar et al., 2011).
Direct electrical stimulation (DES) applied using intracranial electrodes is often considered as the most powerful technique for inferring the critical function of brain regions in humans (Penfield, 1958;Desmurget et al., 2013;see Jonas and Rossion, 2023).The first case drawing attention to the cortical region reviewed here is patient CD who, during focal DES to the right antFG+, was transiently unable to recognize the identity of any famous face picture presented (Jonas et al., 2015).During stimulation, while CD was able to recognize nonface objects presented visually, her behavioral impairment at FIR was clear, massive, and highly reproducible (see videos in Jonas et al., 2015; also Jonas and Rossion, 2021).After stimulation, CD stated that she did not recognize the face identity, as if the picture was shown to her for the first time.She Fig. 4. Human specificity of neural circuits for face identity recognition.A. Ventral view of a human and a macaque brain at relative sizes.The human brain contains an estimated number of 86 billion neurons, about 13.5 times more than a macaque brain (6.3 billion) (Herculano-Houzel, 2016).In humans, the key neural structures for face identity recognition run from the lateral section of the IOG to the midFG and antFG up to the temporal pole.In comparison, the posterior VOTC of the macaque brain is limited to one sulcus, with little gyrification, and no fusiform gyrus (Bryant and Preuss, 2018).B. Inflated segmented brains showing typical locations of posterior face-selective regions in macaques (adapted from Laurent et al., 2023) and humans (adapted from Weiner and Grill-Spector, 2015; with permission under a PMC Creative Commons License).In macaques, these face-selective regions are essentially found in the STS, whereas humans have regions in both the STS and the VOTC.

(caption on next page)
B. Rossion et al. did not report any perceptual change in the structure of the face.
This type of observation was recently reproduced and significantly extended in another case, DN, who also transiently failed to recognize facial identity during right antFG+ stimulation (Volfart et al., 2022;Fig. 6).DN's case was particularly interesting because he truly excelled at FIR outside stimulation and was tested in a large number of trials.Most importantly, beyond his failure at recognizing the identity of a single famous face picture presented upon right antFG+ stimulation, DN was also tested with tasks that did not require a verbal output, quantifying performance in terms of both accuracy rates and response times.Specifically, upon stimulation of the right antFG+ , DN was impaired at pointing out a familiar face among unfamiliar faces and at matching different pictures of the same unfamiliar or familiar identity.However, he had no difficulty at pointing famous names, and naming common objects and famous buildings.As for subject CD, DN never reported visual face-related changes, stating for example: "I don't' know who he is"; "I don't know the 3 faces you showed me", "I had difficulties recognizing her", "I didn't recognize the face immediately" (see Volfart et al., 2022;with videos).
Strikingly, the locations of the critical stimulation sites were very similar across the 2 subjects (posterior part of the antFG+, y Talairach coordinates − 21 and − 30 in DN and CD, respectively; see Fig. 7).In both cases, the critical stimulation sites were located in the antFG+ , just anteriorly to the right FFA, at a location where no fMRI face-selective activations were identified, even at low statistical threshold, because of the strong signal drop out affecting this region (Fig. 7 ).Yet, in both cases, particularly large face-selective electrophysiological activity was recorded on these 2 stimulation sites (Fig. 6 for DN).
Altogether, these two DES case studies show that the face-selective region(s) of the antFG+ , at least in the right hemisphere, is (are) critically involved in FIR.While only two cases have been reported so far, this finding is meaningful, considering that transient category-selective behavioral impairment at FIR (i.e., transient prosopagnosia), not just face distortion (Parvizi et al., 2012;Rangarajan et al., 2014: FIR not tested in these studies), has been objectively reported only in a handful of cases (i.e., only 2 cases beyond these 2, following right IOG/OFA and midFG/FFA stimulation; Jonas et al., 2012;Jonas et al., 2014;and Volfart et al., 2023 respectively).This is in line with the rarity of true cases of prosopagnosia in the neurological population (Sergent and Signoret, 1992;Rossion, 2018;Rossion, 2022a).

A specific role of the antFG in face familiarity or semantics?
Since several sources of evidence (i.e., lesion and fMRI studies) support the view that the ATL is a multimodal semantic region, particularly for person-identity recognition (Joubert et al., 2006;Gainotti, 2007;Busigny et al., 2009;Blank et al., 2014;Rice et al., 2018;Hailstone et al., 2011;Collins et al., 2016), it may be tempting to link the antFG+ to this function.However, ATL regions linked to multimodal person-identity recognition in fMRI are generally located close to the temporal pole, and therefore anterior to the antFG+ (for a review of coordinates of lesions and fMRI activations related to person-identity recognition, see Volfart et al., 2022).Most importantly, the well-documented progressive inability to recognize faces of familiar people in cases of frontotemporal dementia with a right hemispheric dominance (e.g., Barbarotto et al., 1995;Evans et al., 1995;Gainotti et al., 2008;Gentileschi et al., 2001;Gorno-Tempini et al., 2004;Joubert et al., 2006;Busigny et al., 2009; for a review, see Gainotti, 2007) usually concern regions of the temporal pole, anterior to and sparing the antFG+.This is consistent with the finding that DES of the right antFG+ in subject DN (discussed above, Section 6) specifically affected identity recognition of faces, not written names (Volfart et al., 2022;Fig. 6).Admittedly, the latter function appears to depend more on the left than the right ATL (Ranieri et al., 2015;Piccininni et al., 2023).Moreover, one cannot exclude that recognition of identity from the voice which is much less efficient than FIR in humans, requires longer integration time, and is particularly variable and difficult to assess (Hanley et al., 1998;Barsics, 2014; see also Gainotti, 2011) would be impaired by DES to the (right) antFG+ .However, voice (identity) recognition in humans appears to depend essentially on regions of the superior temporal gyrus/sulcus rather than the VOTC (Maguinness et al., 2018;Roswandowitz et al., 2018).Finally, against the view of the antFG+ as being part of a cross-modal semantic region, a recent fMRI connectivity study showed that this anatomically defined region is preferentially connected to posterior face-selective regions (OFA and FFA), as expected from a domain-specific region, rather than being widely connected with different modality-specific regions (Persichetti et al., 2021).In summary, the antFG+ does not appear to correspond to a multimodal semantic region in the VOTC.
Another issue is whether, independently of semantic associations, the antFG+ could be a region particularly sensitive to familiar as compared to unfamiliar faces, in line with the hypothetical role of this region as a seat of so-called 'Facial Recognition Units' (FRU, Davies--Thompson et al., 2014, see Fig. 3B)?In apparent support of this view are early neuroimaging studies with PET, which, as mentioned above, do not suffer from the kind of artifacts found in fMRI.In the study of Kuskowski and Pardo (1999), explicit encoding of unfamiliar face identities in memory for subsequent recall activated the antFG (bilaterally).However, the activation was found in comparison with passive viewing of scrambled images of faces, so that there is no evidence that it was due to the memory task per se.Indeed, passive viewing ('face watching') of unfamiliar faces also activated the same region at the same significance level and, if anything, the contrast in neural activation between the memory task and passive viewing of faces was strongest in the midFG.Two PET studies also identified differential activations for familiar and unfamiliar faces in the antFG (Wiser et al., 2000;Rossion et al., 2001).In the latter study, the right antFG even evoked a clear-cut (i.e., categorical) difference between pictures of familiar and unfamiliar faces in an orthogonal task, with a focus of activation corresponding strikingly to the site of stimulation evoking transient prosopagnosia in the two cases Fig. 5. Identifying and quantifying face-selective activity in VOTC with fast periodic visual stimulation (FPVS) and SEEG. A. In this FPVS (or frequency-tagging) paradigm, natural images of nonface objects are presented at a rate of six images per second (6 Hz) with highly variable face images presented every five stimuli (see green dotted square) (Rossion et al., 2015).Neural activity common to faces and nonface objects is reflected in the SEEG signals at 6 Hz and harmonics, while face-selective activity is reflected at 1.2 Hz (i.e., 6/5) and harmonics.In SEEG, each stimulation sequence lasts for 70 s (3 s showed here).Over the past 10 years, this paradigm has been used and validated in about 30 scalp EEG studies (e.g., Retter and Rossion, 2016;Quek et al., 2018).B. Left, significant face-selective contacts identified across individual brains (here, N = 121 brains) are displayed (in red) over a template brain in Talairach space (Colin27).Empty circles are contacts in which SEEG signal was recorded but that are not face-selective.Right, map of the local proportions of face-selective relative to the number of recorded contacts computed in 12 × 12 mm voxels (for x and y dimensions) in Talairach space and projected to the cortical surface.The dotted contour line shows the location of the antFG+ region over the right hemisphere.Only local proportions significantly above zero (p < 0.01, percentile bootstrap) are displayed.C. Left, face-selective amplitudes in face-selective contacts shown for each anatomical region (i.e., as defined in the individual native anatomy) and separately for the left and right hemispheres (LH and RH, respectively).While face-selective amplitude is highest in the lateral section of the midFG (latFG), i.e., the FFA, especially in the right hemisphere, amplitudes in the antFG+ regions are in the same range as in the IOG.Right, face-selective amplitude computed in 12x12mm voxels and projected to the cortical surface.Abbreviations: VMO: ventro-medial occipital cortex; IOG: inferior occipital gyrus; medFG: medial fusiform gyrus and collateral sulcus; latFG: lateral fusiform gyrus and occipito-temporal sulcus; MTG: middle temporal gyrus; antCOS: anterior collateral sulcus; antOTS: anterior occipito-temporal sulcus; antFG: anterior fusiform gyrus; antMTG: anterior middle temporal gyrus; TP: temporal pole; AMG: amygdala; HIP: hippocampus.mentioned above (see Fig. 6 in Jonas et al., 2015).Importantly, this functional difference between familiar and unfamiliar faces was clearly distinct from (i.e., anterior to) the right FFA in the midFG as defined independently in the same group of subjects (Rossion et al., 2003).While this finding reinforces the important role of the (right) antFG+ in human face recognition, it is important to mention that a similar categorical difference between familiar and unfamiliar faces was also found in the posteriorly located right FFA and OFA in that experiment (Rossion et al.,  (Volfart et al., 2022, with permission).A. Anatomical location of the stimulated contacts (electrode TM) in sagittal and coronal MRI slices.The effective sites inducing a FIR impairment are indicated (TM1-TM2, TM4-TM5 and TM5-TM6).AntCOS, anterior part of the collateral sulcus; AntFG, anterior part of the fusiform gyrus; AntOTS, anterior part of the occipito-temporal sulcus; HIP, hippocampus.B. Face-selective intracerebral responses (as in the paradigm displayed on Fig. 5A) recorded on electrode TM.These responses were determined by summing segments of the EEG spectrum centered on the face presentation frequency and its harmonics (i.e.,1.2, 2.4, 3.6, and 4.8 Hz).The 0 mark corresponds to the face presentation frequency.Note that the most effective contact (TM5) for eliciting transient prosopagnosia recorded the largest face-selective activity, despite the lack of fMRI face-selective responses recorded in this region (see Fig. 7).C. Stimuli and procedure of the famous pointing tasks (face and name) and simultaneous unfamiliar face matching tasks performed during stimulation sessions on electrode TM contacts.During stimulation of the effective sites in the right antFG+, DN was impaired at pointing famous faces among unfamiliar faces, matching unfamiliar faces but not at pointing famous names among unfamiliar names.
2003).More generally, significant differences in neuroimaging signal between familiar and unfamiliar faces have been largely inconsistent across neuroimaging studies (PET and especially fMRI), which reported such differences in widely distributed regions in the brain (temporal, parietal and frontal lobes) and all along the temporal lobe (posterior VOTC especially in the latFG, and ATL) (e.g., Sergent et al., 1992;Von Der Heide et al., 2013;Gobbini and Haxby, 2007;Natu and O'Toole, 2011;Ramon et al., 2015;Weibert et al., 2016).
In addition, PET studies as well as intracerebral recordings and stimulation indicate a key role of the antFG+ in dealing with unfamiliar faces, which are not associated with semantic information, affect or verbal labels, but must be recognized based on visual cues only.In intracerebral recordings for instance, a FPVS paradigm designed to record responses reflecting unfamiliar face identity discrimination (Fig. 8; Liu-Shuang et al., 2014;Rossion et al., 2020), elicits neural activity in strip of cortex stretching along the postero-anterior axis of the VOTC from the IOG, through the midFG and up to the antFG+ , with a right predominance (up to 3 cm from the midFG in the ATL; y Talairach coordinates from − 40 to − 7; Jacques et al., 2020;Fig. 8).Moreover, in subject DN, DES to the right antFG+ region impaired the ability to match simultaneously presented unfamiliar faces for their identity (Fig. 6, Volfart et al., 2022), with the critical contacts recording the largest unfamiliar face discrimination response among all recorded electrodes implanted in the subject's brain.
In summary, while the antFG+ , particularly in the right hemisphere, is critically involved in FIR, there is no clear evidence that its role would be specifically dedicated to explicit encoding or retrieval of face identities, or to the storage of previously encountered faces (i.e., familiar faces), or their direct association with semantic and biographic information.

The Anterior Fusiform Gyrus: an interface between visual perception and memory or just another cortical memory node?
As reviewed above, while the role of the antFG+ is still misunderstood, there is no evidence showing that this region acts as an interface between putative posterior perceptual and anterior (semantic) memory stages.Rather, it appears to share a number of properties with posterior face-selective regions, in the IOG and midFG: high face-selectivity (Jonas et al., 2016;Jacques et al., 2022), sensitivity to individuation of faces based on visual characteristics (Jacques et al., 2020), and criticalness for FIR (Jonas and Rossion, 2021).Two differences should be mentioned, however.
First, as noted initially in the intracerebral recording study of Jonas et al. (2016), the antFG+ presents with a substantial increase in the proportion of face-selective recording contacts with no response Fig. 7.The right antFG+ is critical for FIR, as shown by DES studies.A. Subject CD (Jonas et al., 2015).B. Subject DN (Volfart et al., 2022).In both studies, the left and middle panels show fMRI face-selective activations in the right VOTC (axial and sagittal slices) with critical SEEG electrodes/contacts in the right antFG+ (electrode F in panel A, electrode TM in panel B), inducing a transient FIR impairment when stimulated, shown in red.These critical electrodes are located anteriorly to the most anterior fMRI face-selective activation identified in the midFG within an area with extremely low fMRI signal or SNR (right panel).whatsoever for nonface objects, i.e., face-exclusive responses (Fig. 9; Jonas et al., 2016;Jacques et al., 2022).This indicates an increase in the level of selectivity from populations of neurons in this region as compared to posterior face-selective regions.
Second, the pattern of anatomico-functional connectivity of the antFG+ is likely to differ from posterior (and anterior) face-selective regions.As hypothesized early on based on lesion studies (Rossion et al., 2003;Rossion, 2008), there is now substantial evidence that all posterior face-selective clusters, including the FFA ('pFus-faces' and 'mFus-faces'; see Fig. 1), hold direct connections with early retinotopically-organized visual areas (Finzi et al., 2021; see also Weiner et al., 2016).However, this might not be the case for the antFG+ (Jonas and Rossion, 2021).This may explain why DES to the OFA or FFA(s) may lead to face distortion or feature replacement of the currently experienced visual stimulus (Parvizi et al., 2012;Jonas et al., 2012;Rangarajan et al., 2014;Jonas et al., 2018;Volfart et al., 2023), while the same stimulation to the (right) antFG+ does not lead to such perceptual changes, or even any reported changes in perceptual experience (for subjects DN and CD described above; see Jonas et al., 2015;Volfart et al., 2022).In contrast, in addition to their category-specific impairment during the task, the latter two cases specifically failed to even remember the images that were presented during stimulation (independently of the visual category: faces, objects, names, and the task accuracy), as if this brief episode was not even registered in their memory.Since the antFG+ and the medial temporal lobe episodic memory system connect directly via the rhinal cortex and the hippocampus, as shown by resting-state fMRI, Diffusion tensor imaging and intracranial cortico-cortical evoked potentials studies (Kahn et al., 2008;Catenoix et al., 2011;Libby et al., 2012;Zhang et al., 2016), we speculate that DES to the antFG+ disrupts information flow between this region and the medial temporal lobe, transiently preventing episodic memory encoding (see Jonas and Rossion, 2021).Are these differences between the antFG+ and posterior faceselective regions sufficient to justify a (sharp) functional distinction between perception and memory, as advocated by standard neurofunctional models of human face recognition?We do not think so, for a number of reasons.First, because such a distinction is at odd with the evidence reviewed above in Section 7, indicating a contribution of the antFG+ in matching and discriminating pictures of unfamiliar faces, even implicitly (FPVS-iEEG measures) and when images are presented side-by-side (behavioral measures).Second, because the influential distinction between cases of apperceptive and associative prosopagnosia  Rossion et al., 2020).Face images were presented at a rate of six stimuli per second (6 Hz) in sequences of 65 s.In each sequence, a base face (here ID1) is randomly selected and repeated throughout the sequence.At a fixed interval of every 5th base face (i.e., at the frequency of 6 Hz/5 = 1.2 Hz), a different unfamiliar facial identity is presented (e.g., ID2, ID3, etc.).In a given sequence, faces are either presented all upright or all in the inverted orientation.B. Ventral maps of the local proportion of contacts significant in the upright (left, UP contacts) and inverted (middle, INV contacts) conditions relative to number of recorded contacts.Black solid contours outline proportions significantly above zero at p < 0.01.Right column: to isolate high-level face individual discrimination responses, the local proportion of INV contacts is subtracted to the proportion of UP contacts.Only significant differences of proportions at p < 0.01 are displayed.This map highlights the significant unfamiliar FID responses in the antFG+ , up to 3 cm more anterior than the midFG where the mFus/FFA is typically reported.(Hécaen, 1981;De Renzi, 1986;Barton, 2008;Davies-Thompson et al., 2014; see also Damasio et al., 1990) is not based on convincing evidence.Indeed, all reported cases of prosopagnosia with apparent intact perception of face identities ('associative prosopagnosia') are either substantially slowed down when required to perform simple tasks requiring to match side-by-side pictures of unfamiliar faces for their identity (Newcombe, 1979;Christen et al., 1985;Davidoff and Landis 1990;Delvenne et al., 2004; see also Farah, 1990), indicative of an   Rossion, 2022a).A. In a standard theoretical framework, hierarchical perceptual processes leading to a series of face representations decoded from the sensory input take place before (familiar) faces can match stored representations of these face identities in memory.In this framework, there is a clear-cut distinction between perception and memory for faces, with cases of either apperceptive or associative prosopagnosia potentially observed in neuropsychology.B. In the proposed alternative framework, low-level sensory inputs that successfully match with cortical memories of faces built from experience in the VOTC lead to an initial (holistic) recognition of the stimulus as a face.The phenomenological experience associated with this successful recognition is called 'perception' (which is always conscious).As such, there is no dissociation between perception and memory.Within a few tens of millisecond, between 100-200 ms following stimulus onset, the initially coarse percept is rapidly, and holistically (i.e., without part-based decomposition), refined within the same system.Note that in this alternative framework, for a neurotypical human adult, both familiar and unfamiliar face inputs match category-selective memories in the VOTC.The successful matching, or triggering, of these memories through inputs conveyed by white matter tracts, constitutes the function.There is no view-invariant representation independently of (multimodal) semantic memories, which provide the well-known superiority for generalization of familiar face identity across views.
B. Rossion et al. perceptual impairment, or they present with a multimodal semantic impairment involving person recognition usually following (right) ATL damage or neurodegeneration (e.g., Sergent and Poncet, 1990;Evans et al., 1995;Busigny et al., 2009;Schroeger et al., 2022;see Gainotti, 2010;Gainotti, 2013).Finally, since it implies at least two full representations of the same familiar face identity: one at the apex of a series of perceptual stages, and one stored in memory (Fig. 10A), the (neuro) functional distinction between face perception and memory/recognition stages is not parsimonious.
To account for these issues and for human face identity recognition in general, we propose an alternative theoretical framework (see also Rossion, 2022a) inspired by the theorical constructs of indirect perception (Helmholtz, 1866;Gregory, 1966; see also Purves et al., 2015) and network memory (Fuster, 1997).According to this novel framework, the different face-selective regions in the VOTC, including in the antFG+ , do not reflect a series of (hierarchical) computational stages to build visual representations (e.g., Marr, 1982;Ullman, 1980;DiCarlo et al., 2012; see also Freiwald, 2020;Hesse and Tsao, 2020) but merely constitute cortical memories of faces.That is, populations of neurons in the VOTCfrom the IOG to the temporal polethat have learned through experience to fire selectively to arguably the richest and most commonly encountered category of stimuli in the human (visual) environment, i.e., faces, clustering together at various VOTC locations.These neural clusters would be gradually formed during development by a combination of simple hebbian learning rules (see Fuster, 1997).Due to cytoarchitectonics, connectivity and retinotopy constraints, they would not appear at random locations but with a certain degree of consistency across individual brains (e.g., in the lateral rather than the medial portion of the fusiform gyrus; Weiner et al., 2017).However, the size, precise location and most importantly the number of these face-selective clusters can vary substantially across individual human brains (Gao et al., 2022), reflecting variable biological (i.e., genetic) constraints and differential experiences with faces in the environment throughout development.
As advocated previously (Rossion, 2022a;Rossion, 2022c), this framework has the advantage of being parsimonious: it does not require a series of hierarchical computational visual representations and separate representations to match for previously stored faces in memory.A visual stimulus would be recognized as a face simply when low-level (non-categorical) inputs from early visual cortices successfully trigger activity in these cortical face memories (Fig. 10).In this framework, perception (of faces) is not dissociable from memory, and is defined as the subjective (conscious) experience of recognition.Hence, even severely degraded inputs or facelike objects that trigger activity in the network of regions are recognized as faces (Hadjikhani et al., 2009;Wardle et al., 2020), this (generic face) recognition process being accomplished between 100 ms and 200 ms following stimulus onset.
An outstanding issue is to determine how facial identities can be reliably distinguished across neural activity in these clusters.According to the proposal, representations of facial identities are not located in a given region, but thought to be distributed among the many faceselective clusters that interact dynamically through reentrant connections (Fig. 11;Rossion, 2022c).Rather than constituting hierarchical processing stages, these clusters gradually increase/decrease in sensitivity to semantic/visual (configurations of) features along the postero-anterior axis of the VOTC, depending on their specific patterns of extrinsic connectivity (Fig. 11).Contrary to a view with definite representational/computational stages, this visuo-semantic gradient is compatible with the variability in number of face-selective clusters across individual brains (Gao et al., 2022).

Conclusions
Unlike the ghost in the machine of Gilbert Ryle's Concept of Mind (1949), the hominoid-specific antFG+ is a real physical entity in the human cortical face "machinery", this region having been overlooked due to methodological limitation of fMRI but undoubtedly holding critical face-selective clusters of neurons as revealed by intracerebral electrophysiological recordings and stimulation.Hence, contrary to a widely held view in human face recognition research (Fig. 3), there is no gap between posterior and anterior VOTC face-selective clusters, often conceptualized as belonging to perceptual processes vs. semantic/biographic memories respectively.Rather than a functionally unique region dedicated to either perception or memory or the interface between the two, the antFG+ appears to be "just another cortical memory node" in a neurofunctional postero-anterior gradient, with clusters in the right hemisphere being particularly critical for FIR.The study of this region in concert with reentrant connected face-selective clusters of the VOTC should help reformulating/reconceptualizing face recognition, arguably the ultimate recognition function for the human brain.(Rossion, 2022c).Through a gradient of sensitivity from visual to semantic features along the postero-anterior axis of the VOTC, different facial identities could be represented by partly overlapping groups of neurons (e.g., cortical minicolumns; see Mountcastle, 1997) within these clusters.As a consequence, the features defining the overlap between these groups of neurons are submitted to the same visual-semantic gradient.Such gradient de facto yields increased invariance across physical variations in the postero-anterior direction.Note that while five clusters of roughly similar sizes are represented here, their number and spatial extent vary across individual brains (Gao et al., 2022).The antFG+ is located roughly in the middle of this gradient of idiosyncratic cortical face memories.

Fig. 1 .
Fig. 1.The human cortical face network .A. Inflated ventral views of the human VOTC (from Grill-Spector et al., 2017, left, and Yovel and Freiwald, 2013, right, with permission) showing the typical location of face-selective clusters of voxels as measured using fMRI.Note the lack of measured faceselective activity in the anterior half of the temporal lobe, beyond the region labelled mFus-faces/FFA-2.The white star indicates the approximate anatomical location of the anterior fusiform gyrus and surrounding sulci (anterior occipito-temporal sulcus and anterior collateral sulcus).B. View of the VOTC showing both the STS face network and the VOTC face network.The white star indicates the location of the gap between mFus-faces and a face-selective region in the ATL identified in a few studies (ATL-faces) (adapted from Weiner & Grill-Spector, 2015).C. Estimated white-matter tracts connecting early retinotopic visual areas (V1-V2) to face-selective regions, including the anterior temporal regions (AT), seemingly bypassing the location of the anterior fusiform region (white star) (from Grill-Spector et al., 2017, with permission).

Fig. 6 .
Fig. 6.Stimulation in the right antFG+ in subject DN induced a transient category-selective FIR impairment, i.e., prosopagnosia (Volfart et al., 2022, with permission).A. Anatomical location of the stimulated contacts (electrode TM) in sagittal and coronal MRI slices.The effective sites inducing a FIR impairment are indicated (TM1-TM2, TM4-TM5 and TM5-TM6).AntCOS, anterior part of the collateral sulcus; AntFG, anterior part of the fusiform gyrus; AntOTS, anterior part of the occipito-temporal sulcus; HIP, hippocampus.B. Face-selective intracerebral responses (as in the paradigm displayed on Fig. 5A) recorded on electrode TM.These responses were determined by summing segments of the EEG spectrum centered on the face presentation frequency and its harmonics (i.e.,1.2, 2.4, 3.6, and 4.8 Hz).The 0 mark corresponds to the face presentation frequency.Note that the most effective contact (TM5) for eliciting transient prosopagnosia recorded the largest face-selective activity, despite the lack of fMRI face-selective responses recorded in this region (see Fig. 7).C. Stimuli and procedure of the famous pointing tasks (face and name) and simultaneous unfamiliar face matching tasks performed during stimulation sessions on electrode TM contacts.During stimulation of the effective sites in the right antFG+, DN was impaired at pointing famous faces among unfamiliar faces, matching unfamiliar faces but not at pointing famous names among unfamiliar names.

Fig. 8 .
Fig. 8. Unfamiliar face individual discrimination (FID) in the antFG+ (adapted fromJacques et al., 2020, with permission).A. FPVS was used to measure individual discrimination of unfamiliar faces (seeRossion et al., 2020).Face images were presented at a rate of six stimuli per second (6 Hz) in sequences of 65 s.In each sequence, a base face (here ID1) is randomly selected and repeated throughout the sequence.At a fixed interval of every 5th base face (i.e., at the frequency of 6 Hz/5 = 1.2 Hz), a different unfamiliar facial identity is presented (e.g., ID2, ID3, etc.).In a given sequence, faces are either presented all upright or all in the inverted orientation.B. Ventral maps of the local proportion of contacts significant in the upright (left, UP contacts) and inverted (middle, INV contacts) conditions relative to number of recorded contacts.Black solid contours outline proportions significantly above zero at p < 0.01.Right column: to isolate high-level face individual discrimination responses, the local proportion of INV contacts is subtracted to the proportion of UP contacts.Only significant differences of proportions at p < 0.01 are displayed.This map highlights the significant unfamiliar FID responses in the antFG+ , up to 3 cm more anterior than the midFG where the mFus/FFA is typically reported.

Fig. 9 .
Fig. 9. Face-exclusive responses in VOTC measured during FPVS (Jonas et al., 2016; Jacques et al., 2022, see Fig. 5A). A. Examples of face-exclusive responses in individual participants in the right antFG+ region.Face-exclusive responses are characterized by significant activity at the frequency of face stimulation (i.e., 1.2 Hz) and harmonics without a significant response to nonface objects as measured at the base stimulation frequency (6 Hz and harmonics).B. The proportion of face-exclusive contacts increases from posterior to anterior face-selective regions, being highest in the antFG+ region (from Jacques et al., 2022, based on intracerebral recordings in N = 121).

Fig. 10 .
Fig.10.Two different theoretical frameworks for human face (identity) recognition (see alsoRossion, 2022a).A. In a standard theoretical framework, hierarchical perceptual processes leading to a series of face representations decoded from the sensory input take place before (familiar) faces can match stored representations of these face identities in memory.In this framework, there is a clear-cut distinction between perception and memory for faces, with cases of either apperceptive or associative prosopagnosia potentially observed in neuropsychology.B. In the proposed alternative framework, low-level sensory inputs that successfully match with cortical memories of faces built from experience in the VOTC lead to an initial (holistic) recognition of the stimulus as a face.The phenomenological experience associated with this successful recognition is called 'perception' (which is always conscious).As such, there is no dissociation between perception and memory.Within a few tens of millisecond, between 100-200 ms following stimulus onset, the initially coarse percept is rapidly, and holistically (i.e., without part-based decomposition), refined within the same system.Note that in this alternative framework, for a neurotypical human adult, both familiar and unfamiliar face inputs match category-selective memories in the VOTC.The successful matching, or triggering, of these memories through inputs conveyed by white matter tracts, constitutes the function.There is no view-invariant representation independently of (multimodal) semantic memories, which provide the well-known superiority for generalization of familiar face identity across views.

Fig. 11 .
Fig. 11.A tentative account for face identity recognition in the human VOTC, with different face-selective populations of neurons ("clusters") linked by reentrant connections(Rossion, 2022c).Through a gradient of sensitivity from visual to semantic features along the postero-anterior axis of the VOTC, different facial identities could be represented by partly overlapping groups of neurons (e.g., cortical minicolumns; seeMountcastle, 1997) within these clusters.As a consequence, the features defining the overlap between these groups of neurons are submitted to the same visual-semantic gradient.Such gradient de facto yields increased invariance across physical variations in the postero-anterior direction.Note that while five clusters of roughly similar sizes are represented here, their number and spatial extent vary across individual brains(Gao et al., 2022).The antFG+ is located roughly in the middle of this gradient of idiosyncratic cortical face memories.