Learning-related representational changes reveal dissociable integration and separation signatures in the hippocampus and prefrontal cortex

The episodic memory system enables accurate retrieval while maintaining flexibility by representing both specific episodes and generalizations across events. Although theories suggest that the hippocampus (HPC) is dedicated to represent specific episodes while the medial prefrontal cortex (MPFC) generalizes, other accounts posit that HPC can also integrate related memories. Here we use high-resolution functional magnetic resonance imaging in humans to examine how representations of memory elements change to either differentiate or generalize across related events. We show that while posterior HPC and anterior MPFC maintain distinct memories for individual events, anterior HPC and posterior MPFC integrate across memories. Integration is particularly likely for established memories versus those encoded simultaneously, highlighting the greater impact of prior knowledge on new encoding. We also show dissociable coding signatures in ventrolateral PFC, a region previously implicated in interference resolution. These data highlight how memory elements are represented to simultaneously promote generalization across memories and protect from interference.


SUPPLEMENTARY METHODS
Region of interest definition. Anatomical ROIs were used to restrict RSA searchlight analyses. HPC was manually demarcated on a custom template generated from the mean coronal images; MPFC and IFG ROIs were created on the MNI template brain.
Different procedures were used to define HPC and PFC ROIs to account for the high variability in hippocampal anatomy across individuals. These procedures are described in detail below.
A custom coronal template was generated using ANTS 1 . The T2-weighted mean coronal images from a subset of ten participants with canonical hippocampi were selected for template generation. A bilateral HPC ROI was delineated by hand on the coronal template using established guidelines 2 . HPC was further segmented into anterior (head) and posterior (body and tail) subregions for volume analysis using anatomical landmarks as follows. The posterior boundary of the HPC head was the last slice on which the uncal apex was visible 3,4 . The anterior boundary of the HPC tail was the first slice on which the fornix became visibly separated from the HPC 5   (2) integration for intermixed triads, intermixed within -intermixed across; (3) separation for blocked triads, blocked across -blocked within; (4) separation for intermixed triads, intermixed across -intermixed within. Permutation tests were performed as described previously, yielding four p-value maps for each of the three ROIs for each participant.
Each participant's voxelwise p-value maps were converted to z-statistics and the resulting images were warped to the 1.7 mm isotropic MNI template using ANTS 1 .
Simple contrasts functional ROI analysis. We then interrogated individual within- Across-triad change functional ROI analysis. Because our main searchlight analyses indexed significant integration and separation as change for within-relative to acrosstriad comparisons, it is not clear whether items from different triads in the same learning condition became significantly more (or less) similar to one another following learning.
One possibility is that all items from a given condition became more or less similar to one another (depending on the region), with within-triad comparisons changing the most; alternatively, it might be the case that across-triad similarities remained the same, with only the within-triad similarities changing as a function of learning. To assess these possibilities, we performed a follow-up analysis that quantified the degree of change in neural similarity for across-triad comparisons.
Searchlights were run within anatomical HPC, MPFC, and IFG to test for significant increases or decreases in across-triad RS, separately for blocked and intermixed conditions. Within each sphere of the searchlight, a difference statistic representing the change in across-triad similarity from pre-to post-study was calculated and compared against a null distribution. Null distributions were computed by shuffling whether similarity values came from the pre-or post-study scan (within a given comparison) and re-computing the difference statistic for each of 1,000 iterations.
Participant-level voxelwise p-value maps were converted to z-statistics warped to the 1.7 mm isotropic MNI template using ANTS 1 as described previously.
We then interrogated each of the clusters identified as showing one of the four predicted patterns in the main searchlight analysis for signs of change for across-triad comparisons. Average z-statistics representing the change in across-triad similarity for each learning condition separately were extracted and compared with zero (using a twotailed test, allowing for both similarity increases and decreases) using a bootstrapping approach. Correction for multiple comparisons was performed using Bonferroni correction, yielding a critical p-value of 0.005 (0.05/11 clusters).
Control analyses. As encoding order (i.e., whether blocked or intermixed learning occurred first) was counterbalanced across participants, individuals had substantial differences in learning experience that might impact their behavioral performance and/or neural coding. Accordingly, we performed control analyses to test for effects of encoding order on behavioral and neural measures of interest.
Effects of encoding order on behavior. We first interrogated whether behavioral performance was significantly modulated by encoding order. We performed a 3 × 2 × 2 mixed ANOVA with test trial type (AC, AB, BC) and learning condition (blocked, intermixed) as within-participant factors and encoding order as the between participants factor. Memory performance (proportion correct) served as the dependent measure.
Order of blocked versus intermixed learning (i.e., encoding order) did not significantly affect behavior (main effect of order and two-and three-way interactions; all F < 1.91, all p > 0.180).
Effects of encoding order on Δ RS. Average z-statistics were extracted for each participant across every cluster identified in the main searchlight analyses within our a priori anatomical ROIs. These z-statistics represented the degree to which each participant exhibited the effect of interest (i.e., the effect that was significant in that particular cluster). We tested whether these values differed as a function of encoding order using a two-sample t-test. Correction for multiple comparisons was performed as described above, yielding a critical p-value of 0.005. Encoding order did not significantly modulate changes in neural pattern similarities in any region (Bonferroni-corrected α threshold for significance < 0.005; HPC searchlight: all t 24 < 2.06, all p > 0.051; MPFC searchlight: all |t 24 | < 2.63, all p > 0.014; IFG searchlight: all |t 24 | < 2.14, all p > 0.043).
Effects of encoding order on HPC volume-Δ RS relationship. We also performed one-way analyses of covariance (ANCOVA) to interrogate whether the observed relationships between anterior HPC (head) volume and neural similarity measures differed significantly as a function of encoding order. HPC subregion volumes (head, body, and tail) served as the predictor variables; neural similarity (integration in the blocked condition, and separation in the intermixed condition) served as the response.
Encoding order was the grouping variable. Correction for multiple comparisons was