Natural scenes and item co-occurrences.

(A) Stimuli consisted of natural scene photographs from the COCO image database (http://cocodataset.org). The first example image contains a cat and plants, the second contains a cat and a cake, and the third contains a cat and books. (B) We test whether items that share recent patterns of co-occurrence (plant, cake, and book all co-occur with cat) become more similarly represented in the brain. Axes are hypothetical dimensions of neural activity, and proximity of items within this space reflects the similarity of their neural representations.

Methodological approach.

Multiple measures were extracted from searchlights across the brain. (A) The fidelity of item representations was assessed using a voxel-wise encoding model that predicted each voxel’s response to each of the 80 objects. (B) Semantic content was assessed by comparing neural across-item similarity with the similarities contained within a word embedding model. (C) Representational drift corresponded to decreases in within-item neural similarity over time. (D) We calculated sensitivity to recent statistical structure of the stimuli by determining whether the statistical structure in the first half of a session influenced neural across-item similarity in the second half of a session. We subtracted the correlation in the opposite direction as a tight control to ensure our measure reflected a time-dependent influence of recent structure on neural representation.

Item encoding.

(A) A voxel-wise encoding model employing cross-validated ridge regression was used to predict each voxel’s response to the 80 items and to subsequently predict multivoxel activity patterns evoked by individual scenes. Predictions were compared to actual neural responses to determine the fidelity of item representations across the brain. Item encoding was successful within all ROIs (B) and large swaths of cortex (C).

Semantic content.

(A) Neural across-item similarities positively correlated with semantic similarities extracted from a word embedding model in CA1, PHC, and PRC. (B) Item representations were semantic in nature within a subset of temporal, parietal, and cingulate regions that exhibited successful item encoding.

Representational drift across long timescales.

Within regions containing semantic item representations, we asked whether these representations drifted across the 30 sessions (∼8 months). (A) Within-item neural similarity was regressed against number of intervening sessions. Negative slopes indicate representational drift, where item representations gradually changed over time. Circles reflect average item similarity across participants at each time interval, and the solid line indicates the linear fit. Dashed lines indicate fits for individual participants. (B) Drift was significant in all semantic ROIs but not in V1. (C) Only a subset of semantic regions exhibited significant drift in item representations over time.

Influence of recent statistical structure on item representations.

(A) Structural similarity between pairs of items was calculated using a network modularity approach. Items that tended to occur with the same items across trials have a higher likelihood of being assigned to the same module and thus were structurally “similar”. (B) For a pair of items, structural similarity in the first half of a session predicted neural similarity in the second half of a session in PHC, but not in other semantic ROIs or in V1. (C) A whole-brain analysis revealed a significant cluster that overlaps with the PHC ROI (black outline), as well as clusters in right fusiform cortex and bilateral clusters in posterior middle temporal gyrus (not shown).

Relationship between recent and long-term structure.

The similarity between long-term item co-occurrence structure and the structure in the first half of each session predicted the degree of structure-induced change in PHC, indicating that PHC updated its representations in the face of new information.