Event Integration and Temporal Differentiation: How Hierarchical Knowledge Emerges in Hippocampal Subfields through Learning

Behavioral measures of learning

Participants were exposed to repeating sequences of objects superimposed on a colored frame and were asked to make judgments about the pleasantness of the object–color combination for each object. The color of the frame changed every four objects, representing a change in context that was operationalized as an event boundary (Fig. 1A; Heusser et al., 2016, 2018). Previous studies have shown that, for once-presented events, processing time increases at event boundaries, reflecting a cost to transitioning to a new event (Zwaan, 1996; Pettijohn and Radvansky, 2016; Heusser et al., 2018). We investigated whether this marker of event segmentation persists with repetitions, indicating that participants continue to segment repeated events. Single-trial RTs for the pleasantness judgments were entered into mixed-level general linear models (Lo and Andrews, 2015), testing the effects of boundary, repetition, and their interaction. These models revealed a strong effect of repetition, reflecting that participants got significantly faster over repetitions (Fig. 1B), indicating learning (χ(1)2  = 3,499.4, p < 0.0001, AIC reduction 3,497). We also found a main effect of boundary (χ(1)2  = 373.87, p < 0.0001, AIC reduction 372), indicating slower RTs for boundary compared to nonboundary items. We found no interaction of repetition by boundary (χ(1)2  = 0.02, p = 0.89). These results suggest that slower RTs at the boundary persisted through repetitions, as events become well learned (Fig. 1B). To further demonstrate the persistence of slow RTs at the boundary across repetitions, we tested each repetition separately and found a strong effect of Boundary in each repetition (χ(1)2's  > 49.43, p’s < 0.0001; AIC reductions, >47). Thus, boundaries influence RTs for the novel as well as familiar and well-learned events (Zwaan et al., 1995).

After learning each list, we tested temporal memory for adjacent items from the list using a serial recognition memory test. In each trial, participants were cued with an object from the list and asked to select from two options the object that came next in the list (Fig. 1C). We tested temporal memory for three types of pairs: Pos1-2, Pos2-3, and Pos4-1 (the first number represents the event position of the cue, and the second number the event position of the target). As expected from presenting the sequences over five repetitions during learning, participants were highly accurate and confident in their responses, indicating that the lists were well learned (accuracy: Pos1-2 M = 88.24, SD = 11.91; Pos2-3 M = 88.24, SD = 10.33; Pos4-1 M = 88.33, SD = 10.74; high-confidence hits: Pos1-2 M = 81.20, SD = 18.35; Pos2-3 M = 80.96, SD = 18.20; Pos4-1 M = 88.33, SD = 18.60; rate calculated out of total responses) (Between the last repartition of each list and the temporal memory participants observed all items in random order. We expected that given the extensive training that included many repetitions, and since we presented all items regardless of events, this will have minimal influence on behavior. Indeed, we observed high accuracy rates and confidence levels). No difference was observed in accuracy or high-confidence rates for the three memory trials (one-way repeated measures ANOVA with the factor of event position: accuracy, F_(2, 58) = 0.003, p = 0.997; high-confidence, F_(2, 58) = 0.39, p = 0.68).

The event structure of the task was evident in participants’ RTs during retrieval (Swallow et al., 2009; Radvansky and Zacks, 2017). Interestingly, RTs were significantly slower for Pos4-1 trials, when participants had to make a temporal memory judgment for items that were from adjacent events, compared to those drawn from the same event (Pos1-2, Pos2-3; Fig. 1C). Single-trial RTs were entered into a mixed-level general linear model (Lo and Andrews, 2015), which revealed a strong effect of event position (Pos1-2/Pos2-3/Pos4-1; χ(2)2  = 30.23, p < 0.0001, AIC reduction 26.23). Planned pairwise comparisons between specific pairs of positions revealed that RTs in Pos4-1 were slower compared to both Pos1-2 and Pos2-3 (Pos4-1 vs Pos1-2: χ(1)2  = 31.40, p < 0.0001, AIC reduction 29.41; Pos4-1 vs Pos2-3: χ(1)2  = 6.12, p = 0.013, AIC reduction 4.12). This result suggests that the temporal order of memories that bridged an event boundary was less accessible to our participants. This is consistent with previous studies with once-presented events (Ezzyat and Davachi, 2011, 2014; DuBrow and Davachi, 2013, 2014) but extends that work to show that repeated events are still segmented in memory. We also found that RTs for Pos2-3 were longer compared to Pos1-2 (χ(1)2  = 10.42, p = 0.0012, AIC reduction 8.42), suggesting that cueing with boundary items benefits retrieval of items from the same event, consistent with prior work (Swallow et al., 2009; Heusser et al., 2018). We note that participants were instructed to try to recall the target item already when presented with the cue. Thus, one possibility for why we nevertheless observed RT differences is that within event, participants could recall the target already when presented with the cue (and potentially mostly when cued with the boundary), which enabled choosing the target quickly when the target and lure appeared on the screen. Across events, participants might have not been able to recall the target during cue presentation to the same extent. Thus, only when prompted with the target and lure, they had to make a memory decision, which took longer. Together, the behavioral results during list learning and in the following temporal memory test indicate that events were segmented (Zwaan et al., 1995; Hard et al., 2006).

Univariate activation decreases with learning in DG, but not in CA3

To establish that CA3 and DG can be dissociated in human fMRI, we first examined univariate activation (average fMRI BOLD response) in DG and CA3 over the course of learning. To this end, the univariate activation estimates during learning (t statistics resulting from a standard GLM analysis, see Materials and Methods: univariate activation) per each event position and repetition and per each participant, averaged across all voxels in each ROI, were entered into a repeated measures ANOVA with the factors of event position (1 through 4), repetition (1 through 5), ROI (CA3/DG), and hemisphere (right/left). We found a main effect of ROI (F_(1, 29) = 7.69, p = 0.0096, ηp2 = 0.21), indicating that univariate activation differed between CA3 and DG. We also found an interaction of repetition by ROI (F_(1, 29) = 8.013, p = 0.008, ηp2 = 0.22), demonstrating that repetition differentially modulated activity in CA3 versus DG. A marginal effect of event position by hemisphere was also observed (F_(1, 29) = 2.99, p = 0.094, ηp2 = 0.09); all other effects were not significant (F’s_(1, 29) < 2.36, p’s > 0.13).

We followed up on the ROI by repetition interaction by testing the effect of repetition within each ROI, collapsed across hemispheres. Activation levels in each ROI were entered into a repeated measures ANOVA with event position and repetition. We found a significant main effect of repetition in DG (F_(1, 29) = 5.577, p = 0.025, ηp2 = 0.16), but not in CA3 (F_(1, 29) = 0.34, p = 0.56). No effects of event position nor an interaction were observed (F’s_(1, 29) < 2.20, p > 0.14). These results reflect a decrease in univariate activation in DG from the first repetition to the fifth repetition, while CA3 activation levels remained similar and even numerically increased (DG: Rep1: M = 0.082, SD = 0.15, Rep5: M = 0.001, SD = 0.17; CA3: Rep1: M = 0.006, SD = 0.13, Rep5: M = 0.027, SD = 0.16; average across event positions, the decrease in univariate activation in DG was not consistent across repetitions; Table 1). Thus, in our data, CA3 versus DG revealed different activation profiles. We speculate that a lower BOLD signal in DG might reflect sparse activation, consistent with pattern separation (Treves and Rolls, 1994; Leutgeb and Leutgeb, 2007; Yassa and Stark, 2011; Reagh et al., 2017) and suggest it might not evolve linearly with learning. In CA3, activity patterns may converge toward event representation (see below), but with no changes in average activation level.

Different event representations in the left CA3 versus DG

To test our main questions regarding neural representations in CA3 and DG, we computed the similarity between multivoxel activity patterns corresponding to sequential objects across learning repetitions (Fig. 2). To specifically target changes in multivariate patterns corresponding to the learning of event structure, we used the first presentation of each list as a baseline and subtracted the similarity values for the first presentation from those measured during all subsequent presentations (Schapiro et al., 2012; Schlichting et al., 2015). This allowed us to specifically target learning-related changes in pattern similarity while controlling for correlations that might stem from the BOLD signal characteristics and analysis steps, which are identical in all repetitions (Schapiro et al., 2012; Mumford et al., 2014; Schlichting et al., 2015; Methods). Similarity values were computed for three different temporal distances (lag 1, 2, and 3 items), that is, the similarity between an object and the immediately following object (lag 1), or the object presented two or three trials later, with lag 3 being the maximum temporal distance within an event. These similarity values between objects that appeared in the same event (within event) could therefore be compared to similarity values between objects with the same temporal distance but from different events (across events; Fig. 2, Materials and Methods: Item–item RSA in time and context). The similarity values per pair of objects were entered into a mixed-effects linear model that examined the interaction between event (within event/across events), ROI (CA3/DG), and hemisphere (right/left). Hemispheres were treated separately due to previously reported lateralized representational similarity effects in hippocampal subfields (Kyle, 2015; Schlichting et al., 2015; Stokes et al., 2015; Kim et al., 2017; Dimsdale-Zucker et al., 2018; Bein et al., 2020) as well as in hippocampal event representations (DuBrow and Davachi, 2013; Ezzyat and Davachi, 2014; Hsieh et al., 2014; Heusser et al., 2016). The three-way interaction between event, ROI, and hemisphere was significant (χ(1)2  = 12.91, p = 0.0003, AIC reduction 10.91), demonstrating different event representations in CA3 versus DG that were also modulated by hemisphere. Thus, we followed up and examined the interaction of event by ROI within each hemisphere. This analysis revealed a strong interaction of event by ROI in the left hemisphere (χ(1)2  = 19.27, p = 0.00001, AIC reduction 17.27), but not in the right hemisphere (χ(1)2  = 0.43, p = 0.51). Therefore, further analyses were limited to the left hemisphere. Left CA3 and DG similarity values are presented in Figures 3 and 5, where we focused on the fifth repetition.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2.

Item–item similarity in hippocampal subfields based on context and time. Top: we extracted the multivoxel activity patterns per item in the list and computed similarity between items that belonged to the same event (within event similarity) versus items that belonged to different events (across events similarity). We also computed similarity for items that were with temporal distance of one item as well as temporal distance of two or three items. Bottom: patterns were extracted in hippocampal subfields, which were automatically segmented. The image shows segmentation in one participant’s anatomical space.

Event representations in the left CA3

To further characterize the nature of the differences between CA3 and DG, we first focused on the left CA3. Following previous models of CA3 (Hasselmo and Eichenbaum, 2005; Howard et al., 2005; Kesner and Rolls, 2015), we propose that contextual stability within the same event together with a shift to a different activity pattern between events can be a mechanism by which attractor dynamics lead to integrated event representations in CA3 (Horner and Burgess, 2014; Davachi and DuBrow, 2015; Horner et al., 2015; Clewett et al., 2019; Liu et al., 2022). Our prior work has shown that the stability of hippocampal representations correlates with temporal memory for event sequences (DuBrow and Davachi, 2014; Ezzyat and Davachi, 2014). Consistent with this, it has also been shown that the magnitude of activation in CA3 during retrieval is related to holistic event retrieval success (Horner et al., 2015; Grande et al., 2019). Thus, we hypothesized that CA3 will integrate representations of information occurring within the same event while separating across different events.

Pattern similarity values for objects with all temporal distances (lag 1, 2, 3), within and across events, were entered into mixed-level models testing the effect of event (within/across), temporal distance (lag 1, 2, 3), and their interaction (first irrespective of repetition). In CA3, we found a main effect of event (within vs across: χ(1)2  = 9.06, p = 0.0026, AIC reduction 7.06; Fig. 3), with higher similarity values within, compared to across events. There was no main effect of temporal distance (χ(1)2  = 2.22, p = 0.13) or interaction effects for event by temporal distance (χ(1)2  = 0.04, p = 0.84). Including repetition (1–5) in the model yielded a significant interaction of event by repetition (χ(1)2  = 4.39, p = 0.036; AIC reduction 2.39). No significant interaction of temporal distance by repetition was found (χ(1)2  = 0.75, p = 0.39) or of event by temporal distance by repetition (χ(1)2  = 0.0025, p = 0.96). Thus, these results revealed that CA3 representations become significantly more similar within versus across adjacent events with learning, irrespective of the temporal distance of the measured items. We next examined the second and fifth repetitions separately, to target patterns early versus late in learning. No main effect of event was found in the second repetition (χ(1)2  = 1.55, p = 0.21). By the fifth repetition, however, when events were well learned, within-event similarity was significantly higher compared to across-event similarity (χ(1)2  = 5.42, p = 0.02, AIC reduction 3.42, Fig. 3; for additional analysis details, see Materials and Methods: How do event representations change through learning within each CA3 and DG?). Importantly, the results remained the same when accounting for univariate activation (within vs across similarity in the fifth repetition: χ(1)2  = 5.31, p = 0.02, AIC reduction 3.31; other results did not change as well). Additionally, a control analysis confirmed that CA3 event representations are not simply explained only by items within an event sharing the same color background alone, as items from different events that shared the same background color (but were not continuous in time) were not represented similarly in CA3 (Rep5: χ(1)2  = 0.33, p = 0.57). Together, these results provide strong evidence for distinct event representations in hippocampal area CA3.

• Download figure
• Open in new tab
• Download powerpoint

Figure 3.

Left: CA3 activity patterns represent events. A, Similarity values (Pearson’s r) within event (orange) and across events (gray) in CA3. Rep5, fifth repetition; Rep1, first repetition. B, Individual participants’ values of the difference between within event and across events similarity. C, The same data as in A but presented per each level of temporal distance. As can be seen, similarity did not differ based on temporal distance in CA3. D, Same as C but using normalized dot product values. N = 30. Data are presented as mean values, and the error bars reflect ±SEM. *p = 0.02, reflecting higher similarity within compared to across events, collapsed across gaps (statistical analysis was conducted by entering similarity values of individual pairs to linear mixed-effects models and conducting model comparisons using chi-square tests; see main text).

CA3 event representations facilitate transitioning between events

If event representations in CA3 distinguish nearby events by allocating them to distinct neuronal ensembles, what consequences, if any, might have on behavioral segmentation? As mentioned above, RTs at event boundaries items are slower compared to nonboundary items, which has previously been interpreted as a cost to transitioning between events (Pettijohn and Radvansky, 2016; Heusser et al., 2018; Zacks, 2020). We asked whether it is easier to transition to an event that is clear and distinct in its representation from the current event and thus might evoke less confusion. This could mean a relatively small cost to transitioning between these events, reflected in only a small RT increase for the boundary item. Alternatively, shifting between very different representations might render the transition to the next event more difficult, resulting in longer RTs for distinctly represented events. To examine these possibilities, we created a metric of neural clustering to index distinct representations of adjacent events by averaging similarity within each of two adjacent events and subtracting from it the similarity across these same events (see Materials and Methods: Are CA3 event representations related to segmentation behavior?). Neural clustering was then regressed with the transition cost: the difference between the RT to the boundary item in the second of these events and the immediately preceding pre-boundary item (the subtraction measure per event pairs closely controls for any baseline differences across lists or participants; see Materials and Methods). As can be seen in Figure 4, we found greater neural clustering in the left CA3 significantly correlated with less cost at the boundary, namely, a smaller RT difference between the pre-boundary and boundary items in Repetition 5 (χ(1)2  = 8.56, p = 0.003, AIC reduction 6.56; β = −0.40, SD = 13.65, reflecting that an increase of 1 SD in neural clustering led to a reduction of 40 ms, on average, in the RT difference between pre-boundary and boundary items). If neural clustering facilitates transitioning between events, it should be driven by the across-event similarity and RTs at the boundary. Indeed, lower across-event similarity (more distinction) significantly correlated with faster RTs at the boundary (χ(1)2  = 5.28, p = 0.022, AIC reduction 3.28). The within-event similarity did not explain RTs at the boundary at the pre-boundary item, nor did across-event similarity explain RTs for pre-boundary items (χ(1)2's  < 1.6, p’s > 0.21). This result suggests that, for well-learned events, distinct event representations in CA3 facilitate transitioning between events.

• Download figure
• Open in new tab
• Download powerpoint

Figure 4.

Neural clustering in the left CA3 correlates with easier event transitions. The difference between CA3 within-event and CA3 across-event similarity (i.e., neural clustering) correlated with an easier transitioning between events, as reflected by a smaller difference between RTs for boundary and pre-boundary items (in ms). Data are from the fifth repetition. Similarity values were z-scored within participant, and the effect is plotted marginalizing over confounding factors (list, event order); thus, data are presented in SD units centered around 0. β = −0.40, ***p = 0.003, results of model comparisons using chi-square test for comparing between mixed-level models.

CA3 representations of different events likely reflect different populations of voxels

So far, we see greater CA3 neural pattern within-event similarity compared to across events and that the extent of neural event clustering facilitates event transition, as measured in response times. We next assessed the extent to which these results are consistent with attractor dynamics, in which different events are represented by different populations of CA3 neurons (Treves and Rolls, 1994; Kesner and Rolls, 2015; Knierim and Neunuebel, 2016). In fMRI, the unit of measure is a voxel so we could ask whether different voxels are engaged for distinct events. While the data reported above shows learning-related emergence of event representations, the correlation measure used (and that is prevalent in RSA studies), while capturing changes in the population of voxels activated for an item, can also be sensitive to changes in overall activation level in only a subgroup of voxels. That is because when activation levels in a subgroup of voxels change, the mean activation level across the entire group of voxels changes correspondingly. Consequentially, the distance from the mean of each voxel changes, and with it, the value of the correlation (note that this is true for Spearman’s correlation as well, not just Pearson’s correlation). Thus, we computed additional measures of representational similarity that are (1) sensitive to the overlap in the population of active voxels versus the level of activation, and (2) sensitive to changes in activation levels, regardless of the population of voxels involved.

To that end, we used two known similarity measures, normalized dot product and the norm difference (Walther et al., 2016), that have been previously used in rodent work to examine similarity in hippocampal subfields (Madar et al., 2019). The normalized dot product is the cosine of the angle between two activity patterns (higher means more similar), normalized by their norms (the sum of squared activation values across all voxels), which makes it robust to changes in levels of activation, even in some voxels. The norm difference is the difference between the norms of two activity patterns; thus, it is only sensitive to differences in the level of activation and is blind to the population of voxels contributing to that difference. Briefly, simulations we conducted show that indeed, changing activation levels in some voxels of nonoverlapping patterns influence Pearson’s correlation and the norm difference, but not the normalized dot product, while changing the overlap between patterns while holding activation levels constant influences both Pearson’s correlation and the normalized dot product, but not the norm difference.

Computing the normalized dot product and norm difference on our fMRI data, we anticipated we would replicate our correlation results in the normalized dot product measure but see no differences between the patterns’ norms. Indeed, the normalized dot product replicated our Pearson’s correlation results, finding higher similarity within compared to across events in the fifth repetition (main effect of event, within vs across event: χ(1)2  = 4.56, p = 0.03, AIC reduction 2.56), with no difference based on temporal distance (main effect of temporal distance: χ(1)2  = 0.017, p = 0.89; event by temporal distance interaction: χ(1)2  = 0.17, p = 0.68; Fig. 3). No effect of event, temporal distance, nor an interaction was observed in the norm difference measure (χ(1)2's  < 0.60, p’s > 0.43). Note that the norm difference does not reflect average activation but the sum of squared activation differences between activity patterns. Thus, it is unlikely that the lack of a difference stems from some voxels reducing activations, while others increasing their activation—this should have led to a major difference in norm difference. Together, the results from the normalized dot product and the norm difference measures provide further evidence that changes in univariate activation level in a subgroup of voxels are unlikely to underly low similarity between events. They are further consistent with similar populations of voxels representing items belonging to the same events, while different populations of voxels represent different events, consistent with attractor dynamics.

Temporal differentiation in the left DG that is sensitive to context

As attractor dynamics in CA3 may pull the representations of items from the same event into a more similar space, a complementary mechanism is needed to distinguish these same sequential representations, to allow knowledge of sequential event details. For sequential events that evolve in time, subevents that appear close in time are the most similar in their temporal information and require disambiguation. In rodents, DG lesions have been shown to impair temporal memory (Morris et al., 2013), and, in humans, a prior study showed increased fMRI BOLD signal in CA3/DG in response to changes in the order of sequences of items (Azab et al., 2014), suggesting sensitivity to temporal order. However, neither of these studies has revealed how items are coded in the DG during the unfolding of events. A recent in vitro study found that temporally similar spike trains inputted to DG tissue resulted in divergent DG activity patterns, providing some initial evidence for pattern separation in time (Madar et al., 2019). Some recent work has provided evidence that distinct representations become revealed across learning for similar trajectories but these experiments could not dissociate time from space (Chanales et al., 2017; Liu et al., 2022; Fernandez et al., 2023).

We investigated the following open questions. (1) Does the DG perform temporal differentiation? (2) If it does, is this sensitive to event structure? As above, similarity values between activity patterns of objects with temporal distances of 1, 2, and 3, both within and across events, were entered into mixed-level models. In the left DG, we found a significant interaction between event and temporal distance (χ(1)2  = 6.37, p = 0.01, AIC reduction 4.37), stemming from lower pattern similarity between objects presented close in time, and only so if they were experienced within the same event (Fig. 5). Temporal distance significantly explained variance within events but not across events (within event: χ(1)2  = 11.18, p = 0.0008, AIC reduction 9.18; across events: χ(1)2  = 0.14, p > 0.71). We refer to this as temporal differentiation and suggest that in the left DG, temporal differentiation is specific to objects appearing within the same event, that is, the same context.

• Download figure
• Open in new tab
• Download powerpoint

Figure 5.

DG activity patterns reflect temporal pattern separation within events. A, Similarity values (Pearson’s correlation) in DG within event (orange) and across events (gray). Rep5, fifth repetition; Rep1, first repetition. B, Individual participants’ temporal pattern separation: the difference in pattern similarity between the temporal distance of one object versus three objects, separately for within versus across events. C, Same as A, but using the normalized dot product values. Data are presented as mean values, and the error bars reflect ±SEM. ⊗ p ≤ 0.01, interaction of event (within vs across) by temporal distance, *p = 0.02, **p = 0.004. Statistical analysis was conducted by entering the similarity values of individual pairs to linear mixed-effects models and conducting model comparisons using chi-square tests. N = 30.

We next examined how temporal differentiation evolved through learning. To that end, pattern similarity values from left DG were entered into a mixed-level model testing the interaction of event, temporal distance, and repetition. This revealed a significant three-way interaction (χ(1)2  = 7.17, p = 0.007, AIC reduction 5.17), suggesting that the modulation of left DG neural pattern similarity by temporal distance and event changed through learning. To better understand these changes, we followed up on this three-way interaction by analyzing the second and fifth repetitions separately. In the second repetition, we found no interaction of event by temporal distance, nor a main effect of temporal distance, either within or across events (interaction: χ(1)2  = 0.02, p = 0.88, main effects of temporal distance within, across events: χ(1)2 ’s < 1.1, p’s > 0.30). However, in the fifth repetition, there was a significant interaction of event by temporal distance (χ(1)2  = 6.27, p = 0.01, AIC reduction 4.27; for additional analysis details, see Materials and Methods: How do event representations change through learning within each CA3 and DG?). Pattern similarity was lower for objects close versus far in time within event, but not across events (within event, main effect of temporal distance: χ(1)2  = 10.21, p = 0.001, AIC reduction 8.21; lag of 1 vs 2: χ(1)2  = 5.59, p = 0.02, AIC reduction 3.59; lag of 1 vs 3: χ(1)2  = 8.11, p = 0.004, AIC reduction 6.11; the same comparisons across events: χ(1)2 ’s < 0.23, p’s > 0.63). Further demonstrating that temporal differentiation only occurred for temporally proximal items within event, similarity was lower between items of temporal distance of 1 within event compared to across events (χ(1)2  = 4.43, p = 0.035, AIC reduction 2.43; the same comparison for lags of 2 and 3 was not significant: χ(1)2 ’s < 1, p’s > 0.30). These results remained the same accounting for univariate activation (event by temporal distance interaction in the fifth repetition: χ(1)2  = 6.22, p = 0.01, AIC diff.: 4.22, other results did not change as well).

Our findings show that DG differentiation is unlikely to happen between items that only share a background color, as the activity patterns of items within event that shared the same background color but had a larger temporal distance (lag of 2 and 3) did not show differentiation. However, to directly examine this, we asked whether, in the fifth repetition, temporal differentiation within event (computed as in the main analysis above) is stronger than differentiation between items that appeared across events with the same background color and event position (e.g., the similarity of items in Event position 1 with the similarity of items in Event position 2 in other events that shared the same background color, and likewise for all items in lag of 1, 2, and 3). While the main analysis that compared across events controlled for objective temporal distance, this across-event comparison reflects larger temporal distance than within events but controls for background color and event position of items. Here as well, there is a significant interaction of event (within vs across same color/position) and “temporal distance” (χ(1)2  = 4.00, p = 0.046, AIC reduction 2.00), and no effect of temporal distance across events (χ(1)2  = 1.32, p = 0.25). This shows that temporal pattern separation (i.e., lower similarity for items that are close in time, as reported above) is specific to within-event similarity, and did not occur between items that shared the same background color but belonged to different events.

To sum up, through learning, items that were close in time and belonged to the same event, and thus potentially shared similar temporal and perceptual information, became more separated in the left DG (O’Reilly and McClelland, 1994; Treves and Rolls, 1994; Leutgeb and Leutgeb, 2007; Yassa and Stark, 2011; Kesner, 2018).

A dynamic inhibition model for temporal pattern separation in DG

Neurophysiological literature showing uniquely high levels of inhibition in DG (Freund and Buzsáki, 1998; Chawla et al., 2005; Coulter and Carlson, 2007; Jinde et al., 2013) that determine the temporal dynamics of neuronal firing (Bartos et al., 2007). Thus, we suggest that inhibitory dynamics may account for the lower similarity between temporally adjacent items (Hasselmo and Wyble, 1997; Myers and Scharfman, 2011; Kesner and Rolls, 2015). Specifically, we propose the DG neurons that are activated for one item (e.g., item n) become inhibited for some time such that they are not active when the immediately following item (n + 1) appears. This results in other neurons representing the n + 1 item and in a low overlap between activity patterns of temporally adjacent items (items n and n + 1). Since inhibition gradually decays in time, for the following item (n + 2), some of the neurons can overcome the lower level of inhibition and are active again, resulting in a slightly higher correlation between items n and n + 2. By item n + 3, inhibition decays enough such that similarity levels are high again and reach the levels of between-event similarity. Between events, this inhibition between DG neurons does not come into play because items can be distinguished based on perceptual features alone, thus DG pattern separation is unnecessary (Fig. 6 and Discussion).

• Download figure
• Open in new tab
• Download powerpoint

Figure 6.

Dynamic inhibition model for temporal differentiation. A, Activated units (yellow circles) trigger inhibition that is maximal in the next trial (n + 1) but then gradually decays (blue to white background). By trial n + 2, partial inhibition leads to some chance for a unit to be activated again (faded yellow). By trial, n + 3 inhibition decays further, such that a unit has a high chance of being active. The curves reflect the resulting correlation between activity patterns. B, Simulation results, Rep5 with a decay rate of 0.2. Similarity values for different temporal distances within event (yellow) and across events (gray). Data are presented as mean values over 500 simulations, and the error bars reflect ±SD. For temporal distance of 1, there are no error bars, as Pearson’s correlation between patterns with no overlapping units is identical regardless of the specific units activated.

We formalized this idea using simulations, to illustrate that the temporally decaying inhibition could, in principle, result in the pattern of the data we observed. We do not offer a comprehensive mechanistic account for temporal differentiation, which would be beyond the scope of the current paper. We increased the level of maximal inhibition (between items n and n + 1) across repetitions but kept the decay in time parameter constant (see Materials and Methods: A dynamic inhibition model for temporal differentiation in DG). Across a range of ROI sizes and rates of inhibition decay, our simulations showed that indeed decaying inhibition results in the pattern of results observed in DG, namely, low similarity for temporally proximal items, with similarity increasing as temporal distance increases (Fig. 6). Conceptually, learning in such a model can be seen as shaping an activity pattern for an item that is differentiated enough from previous items, such that it defeats increasing levels of inhibition.

DG temporal differentiation likely reflects different populations of voxels

The dynamic inhibition model suggests that activation of different populations of voxels might underlie temporal pattern separation. Thus, we used the additional similarity measures we used above (normalized dot product and norm difference) to strengthen the empirical evidence that different populations of voxels, rather than changes in the level of activation (potentially in a subgroup of voxels), underlie the results observed in DG. As above, if the pattern similarity results reflect changes in the activation of distinct voxels and not in the activation level of the same voxels, we predicted that we would replicate our correlation results in the normalized dot product measure, but not in norm differences. Indeed, mean normalized dot product values in the left DG showed a similar pattern to Pearson’s correlation: lower normalized dot product for items that are close in time, only within event, but not across events (Fig. 5). As in Pearson’s correlation, in the fifth repetition, the interaction between event and temporal distance was significant (χ(1)2  = 6.77, p = 0.009, AIC reduction 4.77). Temporal distance significantly explained variance within event, but not across events (within event: χ(1)2  = 10.429, p = 0.001, AIC reduction 8.43; across events: χ(1)2  = 0.33, p > 0.56). Critically, we found no difference based on temporal distance or events or their interaction in the norm difference measure (all p’s > 0.14). Together, the results from the normalized dot product and the norm difference measures provide further evidence that temporally close items within the same event are represented by different populations of voxels in the left DG. These results are consistent with the dynamic inhibition model.

Exploratory analysis: similarity values by event position

As an exploratory analysis, in both CA3 and DG, we examined the stability of representational similarity effects across event positions (1–4) in the fifth repetition. In the left CA3, similarity with the beginning and end of an event (Positions 1 and 4) were numerically higher than similarity with the middle of an event (Fig. 7A). In the left DG, temporal pattern separation between objects that are close in time seemed stable across the different positions in the event (Fig. 7B), consistent with the dynamic inhibition model.

• Download figure
• Open in new tab
• Download powerpoint

Figure 7.

Representational similarity per event position in the fifth repetition (in A). CA3 and (B) DG. Orange: within-event similarity. Gray: across-event similarity. Note that each data is a mean similarity between two event positions. For visualization, each data point is presented twice, each time anchored based on the panel per each event position. The same data points are represented using the same fill color. For example, the similarity between event Positions 1 and 4 is presented in light green, once in the panel anchoring on Position 1 and once in the panel anchoring on Position 4. The similarity values of the first repetition (Rep1) are subtracted from all repetitions. Data are presented as mean values, and the error bars reflect ±SEM. N = 30.

CA3–DG connectivity increases with learning

Because we see that DG temporal differentiation is sensitive to event structure, it remains an open question of how DG is sensitive to the event structure. Our results also show that over learning, neural patterns in CA3 become more stable within events. Thus, one hypothesis is that feedback or communication between CA3 and DG may also increase as learning advances. Indeed, a prominent theoretical model proposes that inhibition in DG, and consequentially pattern separation, is achieved via back projections from CA3 (Scharfman, 2007; Myers and Scharfman, 2011). If pattern separation is mediated via CA3 back projections and was higher in Repetition 5 compared to early learning, we aimed to examine if a concomitant increase in CA3–DG synchronization is evident from the first to the fifth learning repetition. To address that, we used a background connectivity approach. Background connectivity is the correlation between slow fluctuations of BOLD signal in two ROIs, after removing trial-evoked activity using a linear regression (here we used 01–0.035 Hz, slower than the rate of object presentation). It is thought to capture broad, state-level, changes in circuit dynamics (rather than momentary fluctuations) and has been shown to capture different phases of learning in hippocampal subfields (Duncan et al., 2014; Tompary et al., 2015). We found that background connectivity between CA3 and DG did indeed significantly increase from the first to the fifth repetition (first rep: M = 0.34, SD = 0.24; fifth rep: M = 0.49, SD = 0.23, t₍₂₉₎ = 3.07, p = 0.005, Cohen’s d = 0.56; additional repetitions: second rep: M = 0.43, SD = 0.23; third rep: M = 0.41, SD = 0.21; fourth rep: M = 0.45, SD = 0.25).

No event representation in CA1

Although we focused on CA3 versus DG representations during learning of events, for comparison purposes, we report data from CA1 as well. In the left CA1, we found no effect of event or temporal distance, nor an interaction between these effects, or any interaction with repetition (χ(1)2 ’s < 2.21, p’s > 0.13). In the right CA1, we found an interaction of event by temporal distance (χ(1)2  = 4.24, p = 0.039, AIC reduction 2.24), with no event by temporal distance by repetition interaction (χ(1)2  = 1.40, p = 0.24). Further, no main effects of event and temporal distance, nor their interaction with repetition, were significant (χ(1)2 ’s < 1.73, p’s > 0.18). This is consistent with previous human studies that did not find CA1 multivariate pattern representations of spatial context (Stokes et al., 2015) or temporal context (Dimsdale-Zucker et al., 2022).