Shaping of object representations in the human medial temporal lobe based on temporal regularities

Abstract

Regularities are gradually represented in cortex after extensive experience, and yet they can influence behavior after minimal exposure. What kind of representations support such rapid statistical learning? The medial temporal lobe (MTL) can represent information from even a single experience, making it a good candidate system for assisting in initial learning about regularities. We combined anatomical segmentation of the MTL, high-resolution fMRI, and multivariate pattern analysis to identify representations of objects in cortical and hippocampal areas of human MTL, assessing how these representations were shaped by exposure to regularities. Subjects viewed a continuous visual stream containing hidden temporal relationships--pairs of objects that reliably appeared nearby in time. We compared the pattern of blood oxygen level-dependent activity evoked by each object before and after this exposure, and found that perirhinal cortex, parahippocampal cortex, subiculum, CA1, and CA2/CA3/dentate gyrus (CA2/3/DG) encoded regularities by increasing the representational similarity of their constituent objects. Most regions exhibited bidirectional associative shaping, whereas CA2/3/DG represented regularities in a forward-looking predictive manner. These findings suggest that object representations in MTL come to mirror the temporal structure of the environment, supporting rapid and incidental statistical learning.

Figure 1

(A) For each subject, fractals were randomly assigned to be the first or second member of 'strong' pairs or 'weak' pairs. During sequence exposure (middle runs), the order of fractals was generated from these pairs, with the constraint that no pair appeared twice in a row. For strong pairs, the first member was always followed by the second. For weak pairs, the first member was followed by the second on only one-third of trials. To equate the frequency of members from weak pairs, the second member was inserted into the trial sequence on its own on the remaining two-thirds of trials. Trials were presented continuously, with no grouping or segmentation cues to the pair structure other than the temporal regularities. (B) Before and after sequence exposure (first and last run, respectively), the same fractals were presented in a random order. This allowed their representations to be measured while avoiding the concern that the temporal proximity of paired fractals would artificially increase their representational similarity. The same random order was used in the first and last run to equate for any spurious order effects or biases in modeling the BOLD response. For each fractal, the parameter estimates across voxels in each ROI were extracted and arranged into a vector. Pattern similarity was assessed by computing the Pearson correlation of vectors from different fractals. This produced three types of correlations: (1) between members of a strong pair, (2) between members of a weak pair, and (3) between members of different pairs ('shuffled' pairs). In all runs, fractals were presented for 1s, separated by a 1, 3 or 5s ISI. Subjects always performed an orthogonal cover task of detecting grayscale patches appearing randomly on 10% of otherwise colorful fractals. Subjects responded on every trial, indicating whether they saw a grayscale patch or not. See also Figure S1.

Figure 2

Changes in pattern similarity (higher values indicate an increase) from before to after sequence exposure are shown for strong and weak pairs. The baseline (shuffled pairs) reflects the change in correlation for recombinations of fractal images into untrained pairs. Brain images show segmented ROIs on a T2 anatomical scan for a representative subject (R=right, L=left). See text for primary bilateral analyses. (A) MTL cortex: strong vs. shuffled pairs (R, t[16]=3.27, p=.005; L, t[16]=2.27, p=.037), strong vs. weak pairs (R, t[16]=3.30, p=.004; L, t[16]=3.10, p=.007), and weak vs. shuffled pairs (R, t[16]=−1.26, p=.226; L, t[16]=−2.72, p=.015). (B) MTL subregions: strong vs. shuffled pairs (R PHC, t[16]=2.54, p=.022; L PHC, t[16]=1.95, p=.069; R PRC, t[16]=2.55, p=.021; L PRC, t[16]=1.13, p=.276; R ERC, t[16]=1.09, p=.291; L ERC, t<1), strong vs. weak pairs (R PHC, t[16]=2.70, p=.016; L PHC, t[16]=2.83, p=.012; R PRC, t[16]=2.96, p=.009; L PRC, t[16]=1.88, p=.078; R ERC, t<1; L ERC, t[16]=2.25, p=.039), weak vs. shuffled pairs (R PHC, t[16]=−1.05, p=.311; L PHC, t[16]=−2.62, p=.019; R PRC, t[16]=−1.28, p=.220; L PRC, t[16]=−1.79, p=.093; R ERC, t<1; L ERC, t[16]=−3.31, p=.004). (C) Hippocampus: strong vs. shuffled pairs (R, t[16]=3.79, p=.002; L, t[16]=2.20, p=.042), strong vs. weak pairs (R, t[16]=4.27, p<.001; L, t[16]=2.51, p=.023), weak vs. shuffled pairs (R, t[16]=−1.46, p=.165; L, t[16]=−1.78, p=.095). (D) Hippocampus subregions: strong vs. shuffled pairs (R subiculum [SUB], t[16]=2.77, p=.014; L SUB, t[16]=2.66, p=.017; R CA1, t[16]=3.03, p=.008; L CA1, t[16]=2.40, p=.029; R CA2/3/DG, t[16]=3.67, p=.002; L CA2/3/DG, t[16]=2.11, p=.051), strong vs. weak pairs (R SUB, t[16]=2.72, p=.015; L SUB, t[16]=2.58, p=.020; R CA1, t[16]=2.62, p=.019; L CA1, t[16]=3.27, p=.005; R CA2/3/DG, t[16]=4.26, p<.001; L CA2/3/DG, t[16]=2.47, p=.025), weak vs. shuffled pairs (R SUB, t<1; L SUB, t[16]=−1.86, p=.082; R CA1, t<1; L CA1, t[16]=−1.76, p=.098; R CA2/3/DG, t[16]=−2.03, p=.060; L CA2/3/DG, t[16]=−1.27, p=.221). *p<.05; **p<.01; ***p<.001. Error bars denote ± 1 SEM. See also Figure S2.

Figure 3

ROIs for each subject were warped to a common template (green). Searchlights within right and left PHC, left PRC, left SUB, right CA1, and right CA2/3/DG showed a greater increase in pattern similarity for strong vs. weak pairs (p<.001 uncorrected). The lack of a left PRC increase in ROI analyses suggests that noise from other voxels may have swamped a local effect. The lack of searchlight effects in right PRC, right SUB, left CA1, and left CA2/3/DG despite increases in ROI analyses suggests that the underlying representations were distributed beyond the scope of searchlights, that we benefitted from the increased statistical power of ROI analyses and greater voxel sample sizes, and/or that specific locations of local changes were misaligned across subjects. For visualization, differences in pattern similarity for strong vs. weak pairs were assigned to the center voxel of each searchlight and resulting statistical maps were thresholded at p<.01. See also Figure S3 and Table S1.

Figure 4

Asymmetry index. If the first member of a pair (fractal A) elicits the representation of the second member (fractal B) but not vice versa, then the correlation of A post-learning with B pre-learning should be greater than that of B post-learning with A pre-learning. Among the seven ROIs that showed increased pattern similarity for strong pairs, only right CA2/3/DG was reliably asymmetric. *p<.05. Error bars denote ± 1 SEM. See also Figure S4.