Selective responses to faces, scenes, and bodies in the ventral visual pathway of infants

Current Biology 32(2):265–274.e5, doi 10.1016/j.cub.2021.10.064; NIH author manuscript, PMC8792213. Read by researcher R4e (innate against learned) for research batch R4, 5 October 2026. Provenance: papers/base/kosakowski2022_infant_category_selectivity.provenance.json.

What was read

• Read in full: every line of the PMC author manuscript, converted from XML: summary, introduction, results, discussion, STAR Methods, acknowledgements, references, figure captions, key resources table and highlights.
• Not read: the supplement (Tables S1–S3 with the statistics and subject details; Figures S1–S3, including the coil comparison and stimulus statistics); figures as images.
• A defect in the source: the sentence on the anterior temporal lobe (ATL) face region is garbled in the PMC text ("face responses were9Fig. 3c"). Its result is taken from the paper's own summary sentence that follows it.

What they did

• Question: are the adult face area (FFA), scene area (PPA) and body area (EBA) already selective in infancy, or do they take years of experience? Earlier fMRI work found them developing slowly through childhood, but had not tested children under 3; the one earlier awake-infant fMRI study (Deen 2017, n = 6) found adult-like locations but no selectivity.
• Participants: 87 infants recruited in the Boston area (2.1–11.9 months, mean 5.0; 162 visits). Usable data from 52. The main results come from 26 infants scanned with a new 32-channel infant coil ("Coil 2021"), aged 2.1–9.7 months; a second set of 26 was scanned on the older coil used by Deen.
• Data: 514.3 minutes of usable data from the new coil, after very strict motion criteria. The selectivity analysis included 19–20 infants (the text says 19; the methods and figures say 20 datasets, aged 3.0–9.7 months).
• Stimuli: blocks of videos and still images of children's faces, children's bodies (hands and feet), toys, and mountain or pastoral scenes, against a baseline of abstract curvy patterns.
• Analysis: in adult-defined anatomical parcels (enlarged slightly to absorb registration error: an unplanned change; results for the original parcels are in the unread Table S1), the top 5% of voxels preferring the target category over objects were selected in one part of each infant's data, and responses were measured in held-out data. Linear mixed models with age, sex and motion; one-tailed tests.

Main results (verified)

• Each region is selective for the same category as in adults, in the same place.
  • FFA: faces above each of bodies, objects and scenes (all p < 0.001). Selective from the top 1% to 30% of the parcel, about 217.5 mm³.
  • PPA: scenes above each other category (all p < 0.01), though not above baseline. Selective even with 100% of the parcel, at least 1,320 mm³.
  • EBA: bodies above each other category (all p < 0.05), but only for the top 1–5% of voxels, about 131.6 mm³. Most face-responsive voxels in the EBA parcel responded to faces and bodies alike.
  • Specificity: the FFA parcel held face-selective voxels and no scene- or body-selective ones; the PPA held scene-selective voxels only.
• Not every adult region is there yet. Face-selectivity in the ATL was significant; the occipital face and place areas (OFA, OPA), the fusiform body area and retrosplenial cortex were not significantly selective (OFA faces > scenes p = 0.004 only). The authors warn that null results in infant fMRI are suspect.
• Data quality explains the earlier negative result. In matched analyses (n = 39), selectivity appeared in new-coil data but not old-coil data: experiment × condition × region interaction F(6,108) = 2.38, η² = 0.013, p = 0.013.
• Low-level features do not explain it. The FFA's face response exceeded both the objects and the curvy baseline, the stimuli richest in curvature and low spatial frequency. The PPA preferred scenes over objects, the most rectilinear stimuli. One puzzle remained: the PPA responded as much to the curvy baseline as to scenes (p > 0.2).

The authors' interpretation

• Category-selective cortex does not need years of experience. Regions sit in about the same place in every individual, so patches of cortex are predisposed for particular functions; continued development may expand and refine them.
• Against "protomaps": the view that cortex starts as maps of low-level features (retinotopy, spatial frequency, curvature) and that categories emerge slowly with experience. If protomaps plus experience produce these regions, "only a few months of visual experience must be sufficient".
• Against strictly bottom-up, hierarchical emergence: the FFA and PPA were selective while their supposed precursors (OFA, OPA) were not.
• Their speculation: location is set by pre-existing long-range connectivity, including to parietal and frontal cortex. Cited support: connectivity in pre-readers predicts where the visual word form area will appear after they learn to read; congenitally blind people have face-selective auditory and tactile responses in the fusiform face area.
• Open, in their words: five-month-olds already have hundreds of hours of visual experience. Which experience, if any, is necessary, and whether these regions compute what adult regions compute, is unknown.

Limits

• Small samples after exclusion (about 20 infants in the key analysis), strict but data-hungry inclusion, and one-tailed tests.
• Parcels were enlarged after looking at the data; the unenlarged results are in the unread supplement.
• Registration to a functional template, without anatomical images for many infants.
• The youngest infant was 2.1 months old, so the study cannot separate innate structure from fast early learning. Some sample counts disagree between text, methods and figures (19 against 20; n = 29 against 23 for the group maps).

What it means for Amadeus (inference)

• What every human meets builds the same map, fast. Faces, places and bodies are inputs every infant meets, and within months they occupy the same cortical locations in every individual. For the innate-against-learned question, this is the clearest case of a shared organisation present before culture-specific content could matter.
• It still does not settle innate against early-learned. The authors themselves leave open whether a few months of universal input are needed. For the base, the practical reading is the same either way: the organisation for universal inputs belongs in the base, as architecture, prior, or training on what every human meets. Neither reading puts it in the slot.
• Connectivity as the prior. If connectivity fixed before experience decides where a function will land (their speculation, with the reading-area example), "architecture plus learning on universal input" is the right shape for the base. Even reading, a cultural invention, lands in a place prepared in advance.