Kurisutina
Library/Recording one brain

Charest et al. · 2014

Unique semantic space in the brain of each beholder predicts perceived similarity

PNAS 111(40):14565–14570, doi 10.1073/pnas.1402594111; free full text, PMC4209976. Read by the main session for research direction R3, 4 October 2026. Provenance: papers/amadeus/charest2014_semantic_space.provenance.json.

What was read

  • Read in full: the PMC HTML, converted to text: significance statement, abstract, introduction, results, discussion, figure captions (figures are images), Materials and Methods, and the reference list.
  • Not read: the supplementary file (pnas.201402594SI.pdf, 1.7 MB), which holds the ROI definitions, the randomisation tests, within- and between-category analyses, ROI-size effects, and the precision estimates. Claims about those rest on the main text only.

What they did

  • Participants and images: 20 people, in 10 pairs. Each saw 72 images:
    • 18 personally meaningful photos from their own life (their own face, a parent's face, partner, pet, bedroom, house front, keys, vehicle, and so on);
    • the 18 photos of their pair partner;
    • 36 unfamiliar standard images.
  • Scanning: fMRI (3 T, 3 × 3 × 3.75 mm voxels, TR 2 s) on two days, a median of 11.5 days apart.
  • Behaviour: similarity judgments by arranging images on a screen. The method's test–retest reliability is r = 0.81, from a cited study.
  • Analysis: representational dissimilarity matrices (RDMs: 1 minus the pattern correlation for each image pair) in early visual cortex (EVC) and human inferior temporal cortex (hIT), each the 1,200 most visually responsive voxels.
  • Individuation index (i-index): the within-person RDM correlation across days minus the between-person correlation, tested by permuting subject labels.

Main results (verified)

  • Representations are stable across days. Within-person RDM correlation, unfamiliar images: EVC 0.32, hIT 0.42. Personally meaningful images: EVC 0.20, hIT 0.40. All p < 0.0001. Between people the correlations were lower but significant: unfamiliar 0.23 and 0.32, familiar 0.17 and 0.30.
  • Each person's representation is partly unique.
    • i-index for unfamiliar images: EVC 0.09, hIT 0.10 (both p < 0.0001).
    • Personally meaningful images: hIT 0.10 (p < 0.005); EVC 0.03 (not significant).
    • The individually unique component is about 12.5% of the non-noise pattern variance in hIT.
  • Brain predicts the person's own judgments, but only for their own meaningful objects.
    • hIT RDMs correlate with similarity judgments (about 0.22 within and between people). EVC does not (about −0.01).
    • The brain–behaviour i-index was significant only in hIT and only for personally meaningful objects (0.05, p < 0.01). It was not significant for unfamiliar objects (−0.008).
  • The authors' caution: prediction of the idiosyncrasies is "robustly better than chance" but "not very precise". The precision figures are in the unread supplement.

Limits

  • 20 young adults; static photos; one high-level visual region. The idiosyncratic component is small, and its predictive precision is modest.
  • RDMs from fMRI patterns are partly shaped by anatomy (folding). The brain–behaviour link argues that the unique component is functionally relevant, not purely anatomical.
  • The proposed medial temporal lobe (memory) mechanism for the personally meaningful effect is the authors' interpretation, not tested here. They found no expansion or contraction of distances related to familiarity.

What it means for Amadeus (inference)

  • Measured, person-specific semantic structure exists, and it matters. Beyond the shared categorical layout, about an eighth of the reliable pattern variance in high-level visual cortex is unique to the person. That unique part predicts how this person judges similarity, but only for things that are personally meaningful to them. For the slot this points at the right content: the person's own world (their people, places, objects), not generic stimuli.
  • Supports P3 and the scan question, with a size warning. A measured functional variable (representational geometry) carries something the person's generic profile would not. The effect is small and imprecise at fMRI resolution with 72 images over two sessions. A month of recording multiplies sessions and items, which should raise precision. Whether it adds beyond the person's own judgments is not tested here.
  • A ready-made measurement for the slot. Similarity arrangements of the person's own items are cheap behavioural probes of their semantic space, with test–retest r = 0.81. They could anchor the "association graph" of brief 5.1 and serve as labels for neural data. Scans then test whether neural geometry adds beyond the arrangement itself.

This summary is our record of the paper, written after reading the full text and published as written; links into our own repository have been removed.