Hippocampal representation of related and opposing memories develop within distinct, hierarchically organized neural schemas

Abstract

Recent evidence suggests that the hippocampus may integrate overlapping memories into relational representations, or schemas, that link indirectly related events and support flexible memory expression. Here we explored the nature of hippocampal neural population representations for multiple features of events and the locations and contexts in which they occurred. Hippocampal networks developed hierarchical organizations of associated elements of related but separately acquired memories within a context, and distinct organizations for memories where the contexts differentiated object-reward associations. These findings reveal neural mechanisms for the development and organization of relational representations.

Figure 1

(A) Training protocol: Rats initially learn problem set XY, then in a new pairs of contexts, problem sets AB then CD. (B) While learning XY, trials to criteria in Context 0 was strongly correlated with that in Context 00. (C) Following initial XY learning, rats rapidly acquired AB and CD. (D) Performance across all nine days of training. Rats performed above chance on AB by the fifth trial block on the first day of training (Day 1: AB1). In contrast, rats performed above chance on CD on the second trial block on the first day in which those items were introduced (Day 4: CD1). There were 15 trials per block. Error bars are S.E.M.

Figure 2

Peri-event time histograms (PETHs) centered on the onset of item sampling of example cells for the four items presented within each position. Y-axis is the trial average firing rate (Hz; scale at upper left for each cell). Grey shading indicates the minimum sampling period. See also Figure S4 for histological confirmation of recording sites.

Figure 3

(A) Firing rates for all cells during ABCD, sorted for CA3 and CA1 by condition that elicited maximal firing rate. Z-axis is the trial averaged z-normalized firing rate. X-axis sorts trial types by Context, Position, Valence, and Item. Strong item coding is reflected in different firing rates among items within a position. (B) Simultaneously recorded cell ensembles for seven trials within an example session. Each histogram is the population vector composed of z-scored (range −1.95 to 7.02 SD) firing rates during one trial identified by Context (Con), Position (Pos), Item (A,B,C,D) and reward valence (+,−). (C) An example correlation matrix from one session showing correlation coefficients by color code (right scale). (D) The mean correlation coefficients (+SEM) for within- and between-condition item sampling events for each task dimension (see Table 1). For all dimensions except set, the correlation coefficients are higher for within-condition trials than between. IVSPC as defined in Table 1. See Figure S2 for individual rat data. (E) A dendrogram showing that ensembles of the same valence in the same position are most similar, followed by ensembles associated with items of opposing valence within the same position. Positions within the same context were also coded more similarly than positions in the opposing context. See also Figures S1-S3.

Figure 4

During item sampling, position is coded first, followed by item, and finally valence. (A) The ensemble correlation analyses was done using the population firing rates taken from 250ms bins centered around sampling for trials in the same location. Sampling epochs shorter that 1.5s were excluded. IVSPC defined in Table 1. Mean of each rat’s average correlation coefficient is plotted with S.E.M. (N = 5). (B) The average d’ for item (black) and valence (red) and position (grey) for trials in the same position. Color coded bar above graph shows periods in which that dimension was significantly coded.

Figure 5

New items are encoded within an established schema. (A) The z-score firing rate for two cells recorded on the last day of training on the first item set (AB3) and the first day of training on the second item set (CD1) 24 hours later. (B) The mean z-score firing rate recorded on the first and last encounter with each item, as identified on AB3 training. (C) Data from AB3 and CD1 were merged and the median ensemble rates for each item and place conjunction were calculated and population vectors were correlated from AB3 to CD1 to create the 8x8 similarity matrix. (D) The valence, position and context coding that developed over AB training was preserved and extended to training with the new item set. IVSPC defined in Table 1. # p = 0.053, * p < 0.05, ** p < 0.01. See also Figure S5.