Development differentially sculpts receptive fields across human visual cortex

Receptive fields (RFs) processing information in restricted parts of the visual field are a key property of neurons in the visual system. However, how RFs develop in humans is unknown. Using fMRI and population receptive field (pRF) modeling in children and adults, we determined where and how pRFs develop across the ventral visual stream. We find that pRF properties in visual field maps, V1 through VO1, are adult-like by age 5. However, pRF properties in face- and word-selective regions develop into adulthood, increasing the foveal representation and the visual field coverage for faces in the right hemisphere and words in the left hemisphere. Eye-tracking indicates that pRF changes are related to changing fixation patterns on words and faces across development. These findings suggest a link between viewing behavior of faces and words and the differential development of pRFs across visual cortex, potentially due to competition on foveal coverage.

The receptive field, the portion of visual space from which information is processed, is a fundamental characteristic of the visual system. Receptive fields are found from the earliest stages of the visual system in retinal ganglion neurons 1 , to V1 2 , to high-level visual regions [3][4][5][6] including regions involved in face 4,5 and word processing 6 .
Given behavioral differences across children and adults in both low-level (e.g. visual acuity 7 ) and high-level (e.g. face recognition 8 ) visual behaviors reliant on receptive fields, it is possible that receptive fields continue to develop across the entire ventral stream after age 5. However, fundamental questions remain unanswered: (1) Do receptive fields in human visual cortex develop during childhood? (2) If so, what is the nature of the development? (3) What is the relationship between receptive field development and viewing experience? Understanding receptive field development will provide fundamental insight into the most basic computation underlying the function of over 30% of the human brain. With disorders such as dyslexia and autism having been associated with atypical brain processing as well as uncharacteristic fixations patterns 9,10 , understanding the link between receptive field development and viewing experience has broad implications in neuroscience.
High-level visual abilities such as reading and face recognition rely on a series of visual computations across the ventral visual stream 11 : a hierarchy of visual areas beginning with V1 and culminating in ventral temporal cortex (VTC) where face 12 -and word-selective 13 regions supporting face 14 and word-form perception 15 , respectively, are located. Since neurons across the entire ventral visual hierarchy have receptive fields [3][4][5][6]16 and neurons with similar receptive fields are spatially clustered 2 , the population receptive field (pRF) of neurons in each fMRI voxel can be reliably measured 17 . In each of early (V1-V3) and intermediate visual areas (V4-VO1) in the ventral stream, pRFs systematically tile the visual field and are organized topographically across the cortical surface into visual field maps 16,17 . In high-level ventral regions that are involved in reading 6 and face recognition 4,5 , pRFs are large and cover the central visual field, generating an over-representation of the fovea, referred to as a foveal bias 18 , rather than a uniform coverage of the visual field.
Theoretical models make different predictions with regards to the first question: Do pRFs develop across the ventral stream? One possibility is that pRFs across the entire visual stream are early-developed or innate. This hypothesis emerges from research showing that the wiring of the visual system which determines neurons' receptive fields and topographic organization is laid out during embryonic development by molecules that guide axon generation and synaptic formation [19][20][21] . A second possibility is that there is a gradient of development, whereby earlier visual areas develop prior to higher-level regions in the ventral stream. This hypothesis is predicted by empirical findings showing that functional [22][23][24][25][26] and anatomical 27,28 development of face and character-selective regions is protracted compared to earlier regions 29 . A third possibility is that pRFs across the entire ventral stream continue to develop during childhood. This hypothesis is suggested by data illustrating that coarse receptive field properties are instilled via embryonic wiring, but that visual experience is necessary to fine-tune them [19][20][21] , as molecular cues alone cannot specify the precision of adult receptive fields and visual field maps 19,21 .
A second, related question is: What neural changes occur during development?
One possibility is that development of pRFs and visual maps is associated with qualitative changes from childhood to adulthood. For example, perhaps not all visual field maps beyond V1 are fully formed in children. A second possibility is that developmental changes are quantitative, but not qualitative. This possibility predicts a similar functional organization of visual field maps in children and adults, even as pRF properties continue to be fine-tuned throughout development. An influential theoryeccentricity bias 18,30,31 -makes specific predictions regarding face-and characterselective regions. In brief, the eccentricity bias theory suggests that because face and word processing require high visual acuity enabled by foveal vision, foveation on faces and words during development leads to the emergence of face-and character-selective regions on existing cortical foveal representations. One version of this theory further suggests that competition between representation of faces and words on foveal resources during development together with left lateralization of the language system in the brain is what generates the adult left brain lateralization for words and right brain lateralization for faces 30,31 . However, the eccentricity bias theory does not make specific predictions regarding the development of pRFs and foveal bias in face-and characterselective regions. One possibility is that pRFs and foveal bias are innate or earlydeveloped, which sculpts the later development of face-and character-selectivity as reported by prior studies 22,25,31,32 . An alternative possibility is that the foveal bias continues to develop throughout childhood, increasing in the left hemisphere within character-selective regions and in the right hemisphere in face-selective regions, consequently enabling more proficient processing of words and faces, respectively.
Is viewing behavior linked to pRF and visual field map development? A large body of behavioral research has shown that fixation patterns in adults are taskdependent, placing their foveal resources on task-relevant information. For example, during face recognition, adults tend to fixate on the center of faces 18,33 (nose bridge) putting informative features 34,35 at the region with the highest acuity. However, it is unknown if children fixate on faces and words in the same way as adults, or if their viewing patterns develop together with the development of pRFs. If fixation patterns are adult-like in children, even as pRFs develop, it would provide evidence supporting the hypothesis that viewing experience shapes pRFs. However, if fixation patterns change together with pRF development, it would suggest that there is a developmental interplay between pRF formation and viewing experience. In turn, this predicts that in order to scan faces and words like adults, pRFs need to be fully developed.
To address these key questions and elucidate the development of pRFs and visual field maps in the ventral visual stream, we modeled pRFs with functional magnetic resonance imaging (fMRI, see Online Methods) in children (n=26, 5 to 12 years old) and adults (n=26, 22-27 years old). Subjects were scanned as they viewed a sweeping checkerboard bar while fixating on a central stimulus and performing a colorchange task. We modeled the pRF of each voxel in the ventral stream as a 2dimensional Gaussian with a nonlinearity, referred to as compressive spatial summation 4,36 (CSS). CSS improves pRF fits in higher-level visual areas 4,36 . We examined: (i) if there are qualitative differences across age groups in pRF properties and visual field maps, (ii) if there are quantitative differences across age-groups in pRF size, pRF eccentricity, and visual field coverage (VFC) obtained by the collection of pRFs spanning each visual area, and (iii) if developmental effects differ across ventral visual stream regions.
To examine if there is a relationship between pRF development and viewing behavior, a subset of subjects participated in a behavioral experiment outside the scanner on a different day. Here, participants viewed images of faces and words during a recognition task while their fixations were recorded by an eye-tracker. We tested if fixation patterns on faces and words differed between children and adults and if so, whether they were associated with pRF properties measured separately during fMRI.

Early and intermediate ventral visual areas are developed by age 5
All subjects completed pRF mapping. There were no significant differences across age groups in (i) motion during fMRI (adult motion average: 0.7mm±0.33mm, child: 0.89mm±0.2mm; t(39)=1.4, n.s.), (ii) fixation behavior during fMRI (t(30)=1.73, n.s. Fig S1A,B), or (iii) task performance during fMRI (t(14)=1.28, n.s., Fig S1C). To test the goodness-of-fit of the pRF model, we measured the mean variance explained by the model for V1 voxels in each participant and compared across age groups. We matched age groups on the variance explained by the pRF model in V1 voxels by excluding 8 children with the lowest V1 model fits and 3 adults with the highest V1 model fits. This matching resulted in no significant differences across age groups in the percentage variance explained by the pRF model across ventral visual regions (Fig S2C). These quality assurance metrics ensure that any developmental effects that we may find are not due to differences between age-groups in motion, performance during fMRI, pRF model fits, or measurement noise.  Examination of the topographic organization of polar angle and eccentricity maps revealed that these maps were qualitatively similar across age groups (Fig 1, Figs S3-S6, all participants' maps). That is, children, like adults, displayed (i) a series of mirrorreversed polar angle maps (Fig 1A,C) emerging from a hemi-field representation in and around the calcarine sulcus (corresponding to V1) and (ii) two sets of large-scale eccentricity maps, one spanning the occipital cortex, in which eccentricities progressively increase from posterior to anterior, and one in VTC, in which eccentricities progressively increase from lateral to medial (Fig 1B,D).
Using polar angle and eccentricity maps, we successfully defined visual areas V1 through VO1 bilaterally in all 18 children and all 23 adults (Figs 1, S3-S6). The cortical volume of visual field maps was slightly (<5%) smaller in children than adults (Fig S2A), but like adults, over 90% of voxels were driven by the mapping stimulus and could be modeled by a pRF (Fig S2B).
Likewise, there were no significant differences across children and adults in either the slopes (F (1,195) =0.39, n.s., 2-way ANOVA with factors of visual area and age group) or intercepts (F (1,195) =2.98, n.s., 2-way ANOVA) of the pRF size vs. eccentricity line fits in V1-VO1. In children, like adults, pRF size also increased across the visual hierarchy, demonstrated by the progressive steepening of slopes of the size vs. eccentricity line from V1 to VO1 (Fig 2C, Fig S7) and the systematic increase in mean pRF size ascending the visual hierarchy (Fig 2A).
As there were no quantitative differences in pRFs properties across children and adults in V1-VO1, the visual field coverage (VFC) obtained by the collection of pRFs spanning each of these visual field maps was strikingly similar across children and adults (Fig 2D). In each of V1 through VO1, the VFC was largely uniform and spanned a hemi-field in each hemisphere in both children and adults. There was no significant difference in the VFC of V1 through VO1 across development (main effect of age group: F (1,244) =1.76, n.s.). Together, these analyses reveal that past the age of 5, children have adult-like polar angle and eccentricity maps, and adult-like pRF properties and VFC in V1-VO1.

The VFC of face-and word-selective regions develops after age 5
To examine if pRFs in high-level regions develop with age, we next defined faceand word-selective regions in all subjects using an independent localizer experiment ( Fig 3A, Online Methods), and compared across age groups mean pRF size, mean pRF eccentricity, and the VFC of each of these regions. We focus on face-selective responses on the posterior fusiform gyrus (pFus-faces) and word-selective responses in the posterior occipitotemporal sulcus (pOTS-chars; Online Methods) because (i) these regions are proximal to the VO1/VO2 transition in VTC, and (ii) a substantial number of voxels in these regions were modulated by the checkerboard mapping stimulus and therefore could be fit by the pRF model (Fig S2E,F). It is noteworthy that in faceselective pFus-faces and word-selective pOTS-chars children had (i) significantly more voxels that were modulated by the pRF mapping stimulus than adults (Fig S2E, F (1,115) =5.68, P<0.02) and (ii) significantly higher percentage variance explained by the pRF model compared to adults (Fig S2F, F (1,114) =8.24, P<0.005, 2-way ANOVA with factors of ROI and age). In general, the size of these regions was not significantly different across age groups (Fig S2D, F (1,115) =0.44,n.s.), except that pFus-faces was numerically larger in adults than children. This difference in voxel number is even smaller considering children had more voxels driven by the bar stimulus in faceselective regions than adults. Additionally, in these regions, there was no correlation between mean pRF size and age or mean eccentricity and age either when considering all subjects or just children (0.35 > Rs > -0.24, n.s.), justifying the grouping of children into one group. In opposition to preceding visual field maps, we found development of pRF properties in pFus-faces and pOTS-chars that varied across hemispheres and regions.
Specifically, pRF centers become more eccentric in the left hemisphere for pFus-faces and more eccentric in right hemisphere for pOTS-chars (Fig 3B).  As the visual field coverage obtained by the collection of pRFs spanning a region depends on the distribution of pRF sizes and eccentricities, subtle development in mean properties may have a profound effect on visual field coverage (VFC) in face-and wordselective regions. To examine this possibility, we estimated the VFC of face-and word-selective regions in each subject, separately for each hemisphere, and then measured the mean VFC of these regions across participants of an age group. As in V1-VO1 there were qualitative similarities in the VFC of face-and word-selective regions across age groups. In both children and adults, the VFC of these regions exhibited a contralateral preference, a foveal bias, and a greater coverage of the lower than upper visual field (Fig 4A-D), as reported previously in adults 4,6 . That is, in each hemisphere, pRFs of face and word-selective regions covered more prominently the contralateral and central visual field than the ipsilateral or peripheral visual field pFus-faces and pOTS-chars, however, differ in their developmental patterns across hemispheres. Specifically, we find significant changes in the VFC obtained by  (Fig 4D, 4E-right). Despite no significant difference in ROI size between groups (Fig S2D), adult pFus-faces are ~30% larger than children. To test if ROI size influences developmental results, we dilated children's pFus-faces to match the mean adult size, and repeated these analyses. Results remain the same (Fig S8), verifying that between-group differences stem from pRF development.
In addition to developmental changes in the CoM, we find significant developmental increases in the total extent of the VFC. That is, the total area of the VFC spanned by pRFs across bilateral pFus-faces and bilateral pOTS-chars significantly increases by ~7 square degrees of visual angle from childhood to adulthood (main effect of age, F (1,86) =5.64, p<0.03, Fig 4F). Together, these data reveal differential development of the VFC in face-and word-selective regions across hemispheres, and an increase in the total extent of VFC.

Development of viewing patterns mirrors pRF changes in high-level visual regions
Previous work suggests that optimal viewing behavior involves central fixations, as the center of the stimulus is the most informative region for recognition of faces and words. This framework predicts similar fixation in children and adults. However, our finding of development of the VFC in face-and word-selective regions may impact viewing behavior on faces and words, respectively. We hypothesized that if pRF coverage guides natural viewing behavior, the optimal behavior would be to place the VFC, not the fovea, onto the center of stimuli. For children, this predicts fixations that are biased off of the center, resulting in systematic shifts in the viewing of faces and words across children and adults. Specifically, the neural data make three predictions: (i) due to the larger foveal bias in adults, they will show more central fixations than children, (ii) if the VFC in right pFus-faces guides fixation, children's fixations on faces will be more rightward and upward biased than adults, and (iii) if left pOTS-words drives fixations on words, children's fixations on words will be more leftward and upward biased than adults.
We assessed natural viewing of faces and words in a subset of our participants (12 children and 11 adults) in a separate behavioral experiment. Outside the scanner, each participant first viewed a series of images from different categories (including faces and pseudowords) and performed a one-back task. Then, participants completed a surprise, self-paced old-new recognition task during which their eye movements were recorded with an eye-tracker (Online Methods). We then determined if free-viewing fixation patterns followed the predictions of the visual field coverages obtained from the fMRI experiment inside the scanner while participants were fixating. Results show that fixation locations on both face and pseudoword stimuli differed between children and adults. As shown for the example stimuli, adults foveate more centrally within face and pseudoword stimuli, while children fixations are more eccentric across the stimulus expanse (Fig 5A). To quantify differences in fixation patterns across age groups, we measured the region of the image in which adults make most of their fixations by calculating for each face and pseudoword stimulus the central region in which adults made 70% of their fixations. Then, we calculated for each child and each image the proportion of fixations made outside of the adult fixation zone and then derived the mean proportion of such fixations across child participants. Results indicate that children fixate significantly outside of the central adult fixation zone for both face (t(11)=4, p<0.01) and word (t(11)=3.63, p<0.01) stimuli (Fig 5B), whereby about 50% of their fixations are outside the adult central fixation zone, even as they make fewer fixations than adults (Fig S9).
Critically, it is not the case that children make more variable fixations than adults, as they show systematic biases in their fixation patterns. Notably, these biases mirror the asymmetries in visual field coverage of face-and word-selective regions in their dominant hemispheres. As shown for the example stimuli, children tend to bias their fixations towards the upper right side of faces (Fig 5A) which puts the VFC of right pFus-faces, which is biased to the left and lower visual field, in a location where it optimally covers the face. Similarly, children tend to fixate on the leftward aspect of words (Fig 5A), putting the VFC of the left pOTS-chars, which covers the right horizontal visual field, in a place where it optimally covers the word. We quantified this fixation bias by calculating the center-of-mass of fixation densities on each face and pseudoword stimulus separately for adults and children. In The eccentricity bias theory 18,30,31 suggests that foveation on faces and words during natural viewing anchors the processing of these stimuli to regions in VTC representing the fovea. Consistent with this view, in both children and adults, the visual field coverage in face-and word-selective regions is foveally biased, providing a more substantial coverage of the central than peripheral visual field. Future longitudinal research in younger participants will determine whether the over-representation of the central visual field emerges before or together with selectivity to faces or words.
Unpredicted by the eccentricity bias theory, our data show that spatial representations in these high-level regions continue to develop from childhood to adulthood. In fact, both the foveal bias and the overall visual field coverage obtained by pRFs in face and word-selective regions increase from childhood to adulthood. These findings argue against the hypothesis that face or word selectivity develop on top of a mature foveal bias and spatial representation. The expansion of the visual field coverage in face-and word-selective cortex and increased foveal bias may involve proliferation of dendritic arbors and synapses to support the increased pooling of information. Thus, pRF development may be associated with microstructural cortical tissue growth that has been observed in face-and word-selective regions 27 .
Notably, the development of pRF properties and visual field coverage also varied by hemisphere across face-and word-selective regions. Word-selective regions became more foveally biased in the left-hemisphere, where previous research has demonstrated lateralization for word-form processing and reading [38][39][40] . By contrast, face-selective regions became more foveally biased in the right hemisphere where face processing is thought to be lateralized 41,42 . Intriguingly, at the same time, visual field coverage shifted away from the fovea for face and word-selective regions in their nonpreferred hemispheres. This pattern of development has important implications for the theory that reading and face recognition compete for foveal representations 31,43 because it provides striking evidence for a competitive push-pull mechanism in which the foveal over-representation increases in one hemisphere and decreases in the other, in an opposing manner across hemispheres for faces and words. Additionally, the retreat of pRF coverage from the fovea in non-preferred hemispheres mirrors previous Our findings are important not only for elucidating the development of a fundamental computation -spatial processing by receptive fields -in the human ventral stream and showing its relation to viewing patterns, but also for providing an innovative methodology and computational framework for investigating development of computations across cortex more broadly. As receptive fields are a basic hallmark of neurons in sensory cortical systems (e.g. auditory 44,45 or somatosensory 46 cortex), as well as characterize complex cognitive tuning (e.g. to numerosity 47,48 ) our novel approach can be applied to quantitatively examine development of cortical function throughout the brain. Likewise, our findings lay fundamental groundwork towards understanding abnormal cortical processing as well as potential maldevelopment in atypical populations, including developmental prosopagnosia 49 , dyslexia 10 , and autism 9,50 .
In sum, we find that early-developed visual field maps in the human ventral visual stream may provide a neural scaffold that shapes the organization of high-level visual regions and that the development of pRFs in high-level visual areas involved in face and word processing is linked to changing viewing patterns on faces and words. Together, these data suggest that both the spatial window through which a region of cortex processes information and our visual experience of complex stimuli changes from childhood to adulthood.

DATA AVAILABILITY
All code relevant to data analysis for the main findings (Figures 1-5) will be available on github.com/VPNL upon request. Any source data relevant to these analyses will also be made available upon request. The majority of the code used in this study was derived from scripts and functions available through the open-source vistasoft code library: https://github.com/vistalab/vistasoft. and behavioral changes, and our experiments required maintaining central fixation, we could not make measurements on younger children where acquiring such data is unfeasible. A similar range of ages was chosen in adults when most structural and functional development in VTC is thought to be near completion 52,53 . Following data quality thresholds discussed below, 8 children and 3 adults were excluded from further analysis (18 children, 23 adults remain). Participants had normal or corrected-to-normal vision and were screened to have no prior or current psychiatric conditions. All (ii) Children completed the recognition memory task with eye tracking outside the mock scanner on the same day in which they participated in training; adults completed this task after scanning was completed. (iii) All participants participated in an MRI session in which we obtained anatomical MRI brain volumes which were used to register data across sessions and obtain cortical surface reconstructions of each brain. (iv) All participants participated in an fMRI session in which we measured brain responses to stimuli of various categories (referred to as localizer experiment). (v) All subjects participated in an fMRI session composed of four runs of pRF mapping.

Data acquisition
Quantitative magnetic resonance imaging (qMRI): Quantitative MRI measurements are obtained from the protocols set forth in 54  University using a phase-array 32-channel head coil. Functional data for the category localizer were collected with a simultaneous multi-slice EPI sequence with a multiplexing factor 55 of 3 to acquire near whole-brain (48 slices) volumes at TR=1s, TE=30ms. Data were acquired at a resolution of 2.4mm isotropic voxels with one-shot T2*-sensitive gradient echo sequence with slices aligned parallel to the parieto-occipital sulcus. Functional data for retinotopic mapping were of similar resolution and orientation but collected on a 16-channel head coil, TR=2s, acceleration factor of 2, 28 slices. fMRI category localizer experiment: The purpose of this experiment was to identify those voxels whose neural response preferred either faces or words in order to localize face-and word-selective cortex as functional regions of interest. During scanning, subjects completed 3 runs, each 318 s long, of an experiment presenting subjects with stimuli from 5 categories each with two subcategories (Faces: child, adult; Bodies: whole, limbs; Places: corridors, houses; Objects: cars, guitars; Characters: words, numbers) as described previously 23,27,56 . Images of a category were presented in 4 s miniblocks at a rate of 2 Hz and did not repeat across miniblocks or runs. Each category was shown 8 times in a run in counterbalanced order interleaved with blanks. Subjects  17,57 . Eye-tracking and fixation task performance were collected on a subset of children and adults (Fig S1). Fixation performance on subjects was tracked with the Eyelink software (http://www.sr-research.com/). Blinks, labeled by the Eyelink software, were removed from the timecourse data of the recorded eye by scrubbing with a 100ms window on either end of the blink. Fixation data was then plotted for each subject. Only subjects that made fewer than three saccades (2 o in size) during a mapping run were included for analysis. Due to the scanner environment, size of participants' head, and time constraints, not all subjects could be eye-tracked during pRF mapping (eye tracking data was obtained for 25 children and 6 adults). Fixation task performance was also only collected on a subset (8 children, 7 adults) of subjects due to button box malfunction. All subjects, however, were trained on proper fixation technique during the recognition memory task (see Measuring Fixation Patterns below), and all subjects included in the analysis that underwent eye-tracking in the scanner fixated successfully, with no difference between age groups. As a reminder, we also observe no difference in pRF properties or pRF model performance in V1 between children and adults, further suggesting proper fixation performance, as improper fixation significantly impacts pRF size estimates 58 .

Data Analysis
Anatomical data analysis: Both the spoiled-GE and the SEIR scans were processed using the mrQ software package in MATLAB to produce T1-weighted maps 54 . The mrQ analysis pipeline corrects for RF coil bias using SERI-EPI scans, producing accurate proton density (PD) and T1 fits across the brain. The full analysis pipeline and its published description can be found at (https://github.com/mezera/mrQ). An artificial T1weighted anatomy was produced for each subject from these quantitative measures which were used for surface reconstruction and visualization of retinotopic data.
Anatomical images for each subject were segmented through FreeSurfer (https://surfer.nmr.mgh.harvard.edu/), the resultant tissue segmentation was handcorrected for classification errors. Functional data were restricted to the cortical ribbon by growing a 3-voxel thick (1 mm isotropic voxels) ribbon from the gray-white matter boundary. fMRI data analysis: Data were processed and analyzed in MATLAB using mrVista software (http://github.com/vistalab) as in previous publications 23,27 . Functional data were aligned to the artificial T1-weighted volume. Functional data were unsmoothed, always analyzed within the individual subject native brain anatomy space, and were restricted to the cortical ribbon.
Functional data were motion corrected both within and between scans. Any subjects who moved more than 2 voxels within a scan were either excluded from data analysis or invited back for another session, such that children and adults were matched for data quality as shown in Fig S2C. There was no significant difference in motion during scanning between groups (see Results). To ensure there were no group differences between children and adults resulting from differences in data quality, agegroups were matched for the mean percentage variance explained of the pRF model across voxels in V1, resulting in no significant difference in explained variance across all visual field maps (F (1,185) =0.59, n.s.).
Definition of V1-VO1: Maps of pRF phase and eccentricity were projected onto an inflated cortical surface reconstruction for each subject (Figs S3-S6). Borders between retinotopic maps were drawn on the cortical surface down the center of phase transitions occurring at the vertical or horizontal meridian 16,60,61 . V1, hV4 61,62 , and VO1 61,62 were drawn as hemifields representing the contralateral visual field. V2 and V3 were drawn as quarterfields separated by V1, and were later combined to produce a hemifield representation. Individual maps were drawn by JG and independently checked by VN and KGS.

Definition of face-and character-selective functional regions of interest (ROIs):
Statistical contrasts of faces or characters > all other stimuli were thresholded at tvalues > 3 for all subjects, as in our previous work 23,27,56 . Face-selective voxels that responded more strongly to faces than other stimuli and were located in the posterior lateral fusiform gyrus were defined as pFus-faces/FFA1. Character-selective voxels that responded more strongly to pseudowords and number strings than all other stimuli that were located on the posterior occipitotemporal sulcus lateral to pFus-faces were defined as pOTS-chars as in 56 . This region is also defined elsewhere as VWFA1 56,63 using real word stimuli. Given that our region (pOTS-chars) occupies the same anatomical location as VWFA1, we refer to it throughout the manuscript as word-selective cortex for simplicity.
Estimating population receptive fields (pRF): After functional data were transformed to the whole brain anatomy and restricted to the cortical ribbon, a population receptive field model was fit in each voxel 17 . For each voxel, a 2-dimensional Gaussian receptive field is modeled, having a center described by x and y coordinates and a sigma describing the width, and a parameter, g, describing its gain. An additional variable is fit for each voxel describing a compressive summation factor of the product of the stimulus and the Gaussian receptive field to better describe nonlinear summation properties of cortical responses as one ascends the visual hierarchy 36 . A candidate timecourse is produced from this pRF by convolving an HRF with the product of the stimulus movie and the pRF. The variables x, y, and sigma are swept until the variance explained of the voxel's timecourse is maximized by the pRF model. Voxels were only included for subsequent analysis if the variance explained by the pRF model was greater than 5%. Additionally, to ensure the most accurate pRF fits, voxels whose pRF centers were outside the stimulus field (>7 o radial eccentricity) or whose sigma was assigned the model's minimum/floor value (0.21 o ) were excluded from further analysis. pRF size versus eccentricity fits (Fig 2): To evaluate the relationship between a pRF's size and its eccentricity, voxels within an individual's ROI were entered into a linear regression in which each voxel's contribution was weighted by the variance explained of the pRF model. Only voxels with greater than 5% variance explained were included.
The line-of-best fit was derived in each subject for each ROI, and then the slope and intercept of this line was averaged across participants of each age group.
Visual field coverage analyses: To calculate the visual field coverage (VFC) for a given ROI and subject, all voxels in an ROI that contain pRFs with >5% variance explained by the model are included and modeled as a Gaussian with a peak normalized to 1. The VFC is produced at each point by averaging the value across pRFs that cover that point, and then normalizing by the maximum coverage value in that subject. We also implemented a bootstrapping procedure 59 that draws with replacement n-voxels from a subject's ROI of size n, and produces an average VFC from 50 iterations to reduce the effect of outlier voxels. The average VFC from this bootstrapping approach is the VFC used for a given subject's ROI. To produce the average VFC of subjects in each age group (Figs 2 and 4), the VFC is averaged across subjects of an age group. For the VFC of the visual field maps shown in  (Fig 4F), we estimated the bilateral VFC for pFus-faces and pOTSchars in each subject. pRF coverage density was binarized in each subject's ROI (nonzero coverage assigned a value of 1) and the proportion of the visual field covered was multiplied by the total area stimulated by the sweeping bar stimulus (pr 2 , r=7 o ), resulting in the square degrees of visual angle covered by an individual's ROI. We then averaged this across subjects in a group.
Center-of-mass distance from fixation: To quantify the foveal bias observed in face-and character-selective regions, we computed the center-of-mass (CoM) distance of the VFC of each region from the center of the visual field (Fig 4). This was derived by multiplying each coordinate by the normalized coverage density to obtain the center of VFC in a given region within children or adults. This measure was then jackknifed, repeated n times leaving out n-1 subjects on each fold, to produce the bars of standard error.
Measuring fixation patterns during free viewing of face and word stimuli outside the scanner: All subjects completed a recognition memory behavioral experiment while being eye-tracked with an Eyelink 1000 eyetracker (www.sr-research.com) in our eye tracking lab before scanning. Participants were seated, head-fixed using a chin rest and positioned 54cm from a monitor and told to freely view stimuli. The experiment had 3 parts: (1) Encoding: participants viewed images from five categories (child/adult faces, indoor/outdoor scenes, car/guitar objects, word/number characters, whole/limb bodies) and performed a 1-back task, indicating when 2 consecutive images were identical. Recognition memory: immediately followed the fixation training. Here, participants were presented with a surprise recognition task in which images appeared on the screen and for each image they were asked to indicate if it was previously seen during the "encoding" phase or if was a new image. This part was self-paced and the images appeared on the screen until participants made a decision. We report fixation from this phase.
Eye movement analysis: After removing timepoints during which participants blinked, data presented in Fig 5 were analyzed in the following way: Fixation patterns were plotted in a 2-dimensional matrix (768x1024 pixel grid, equal in size to the stimulus presentation screen) and smoothed with a small Gaussian filter (sigma=18.75pixels) for the purpose of averaging data across subjects. Fixation density was normalized by the maximum in each subject, and then averaged for a given stimulus across all participants of an age group. The adult average fixation density was thresholded at 70% overlap for each stimulus and defined as the "adult fixation zone", AFZ. The ratio of individual fixations made inside versus outside this this AFZ was calculated for each child participant and image, and then averaged across participants and stimuli of a given class (e.g. faces). The ratio was defined as (fixation time outside AFZ) / (total fixation time). A value of 1 indicates that all fixations occurred outside the AFZ, and value of 0 indicates that all fixations were within the AFZ. We then calculated if children fixated outside the AFZ significantly higher than chance, chance here being that 30% of fixations would occur outside the AFZ (as it was defined in adults as the 70% overlap contour).
Fixation bias vector analysis: The average fixation density for each face and word stimulus from the visual recognition test was calculated separately for children and adults. We first calculated for each image the center-of-mass of the distribution of adult fixations, similarly to the adult-fixation zone analysis discussed above, finding the center of the zone where 70% of adults fixated. We then calculated the center of mass of child fixations. From this center, a vector was produced pointing towards the center of fixation density on the same stimulus in children. Bias in child fixation vectors (Fig 5C) was quantified using a t-test to determine if vectors for a given stimulus category significantly deviated away from a vector bisecting the quadrant where the visual field coverage exists (for example, the coverage of right pFus-faces in the lower left quadrant) which we term the null quadrant. This procedure tested the hypothesis that children fixate in an optimal manner (e.g., they do not fixate in such a way that would move their limited coverage away from the informative region in the stimulus). It was assumed that 25% of randomly distributed vectors would have angles within the null quadrant (if randomly distributed, 25% of vectors should lie in quadrant spanning a quarter of the visual field), and t-tests were performed to assess if resulting bias vectors significantly deviated from this null.
Statistical Analysis: N-way ANOVAs were run for data presented in Bootstrapping methods were used to produce VFC plots in Fig 2 and 4 to ensure robustness of fits and downweight outlier voxels; this bootstrapping method is described in the section Visual field coverage analyses, above. For Fig 4A-D  Reported volume measurements are the volume of voxels within the map that survive variance-explained thresholding. Children are light gray, adults black. Subjects included are matched for variance explained in V1. Subject numbers are the same as those reported in Figures 1 and 4. (B) The proportion of an ROI that is retinotopically driven above the 5% variance-explained threshold. Children are gray, adults black. (C) The mean percentage of variance explained across ROIs in children (gray) and adults (black) after variance-explained thresholding. (D-F) Same as A-C but for face-and word-selective regions. The r-and l-denote right and left hemisphere. ANOVAs for each were run with grouping variables of ROI and age-group. Main factors are reported, followed by the interaction.

A .
B .  The of number of fixations on faces and pseudowords in children (light colors) and adults (darker color) during the recognition task. Adults make significantly more fixations than children (t(21)=2.7, p<0.02), but they are clustered near the center of the stimulus (Fig 5)