Neural Differentiation Tracks Improved Recall of Competing Memories Following Interleaved Study and Retrieval Practice

Abstract

Selective retrieval of overlapping memories can generate competition. How does the brain adaptively resolve this competition? One possibility is that competing memories are inhibited; in support of this view, numerous studies have found that selective retrieval leads to forgetting of memories that are related to the just-retrieved memory. However, this retrieval-induced forgetting (RIF) effect can be eliminated or even reversed if participants are given opportunities to restudy the materials between retrieval attempts. Here, we outline an explanation for such a reversal, rooted in a neural network model of RIF that predicts representational differentiation when restudy is interleaved with selective retrieval. To test this hypothesis, we measured changes in pattern similarity of the BOLD fMRI signal elicited by related memories after undergoing interleaved competitive retrieval and restudy. Reduced pattern similarity within the hippocampus positively correlated with retrieval-induced facilitation of competing memories. This result is consistent with an adaptive differentiation process that allows individuals to learn to distinguish between once-confusable memories.

Keywords: differentiation; hippocampus; memory; neural network model; pattern similarity; retrieval-induced forgetting.

Figure 1.

Predictions of our neural network model. (a) Here, we depict 2 partially overlapping memory representations—Abner the Ape (large pink circle) and Anton the Ape (large orange circle)—after their initial study. Inhibitory interneurons (not shown) enforce an approximate “set point” on the amount of neural activity in the network (O'Reilly and Munakata 2000). For the purpose of this diagram, we assume that the 5 units receiving the most excitatory input within the upper part of the network (i.e., within the large gray rectangle) are allowed to be strongly active; additional units are allowed to be weakly active. We also assume that there exist sensory units (in the blue zone) associated with each animal that represent the animal's sensory features. These sensory units are activated whenever the relevant animal is presented; activation spreads from the sensory units to the rest of the animal's representation. (b) During selective retrieval practice of Anton (the Rp+ item), the units associated with Anton's representation are strongly activated in memory. Because of the overlap, Abner's (the Rp− item's) representation is also partially activated. (c) In the model, strong activation of Anton's representation triggers further strengthening of connections between these units. Also, weak activation of Abner's representation triggers weakening of connections into Abner's units (from other active units). This competition-dependent weakening of connections (highlighted in red) leaves the Abner representation in a degraded (less-fully-interconnected) state. The decrease in interconnectivity is assumed to make the Abner memory harder to recall if memory were tested at this point in time (i.e., RIF). (d) When Abner is restudied, the 2 units that formerly were shared between Abner and Anton are no longer in the “top 5 most excited units” because of the weakening that took place earlier. As such, they drop out of the representation of Abner. Other units that were not previously activated then take their place via spreading activation, leaving Abner with a full complement of 5 activated units. (e) Learning in the model strengthens the connections (highlighted in green) between these new units and the other Abner units. In the final state of the network, the Abner representation is nearly as strong (i.e., its features are roughly as densely interconnected) as it was at the outset, and now it overlaps less with the representation of Anton. This neural differentiation will result in less competition at retrieval, which should boost recall of Anton above baseline (i.e., revRIF).

Figure 2.

Stimuli. Our 6 animal categories, each containing 8 pictorial exemplars, were assigned to either the Rp (retrieval practice) or Nrp (baseline) condition in a counterbalanced fashion across participants. Proper names beginning with the first letter of the relevant category were randomly assigned to the individual pictures, which were also randomly divided into an A-set and a B-set. Image–name pairs randomly assigned to the A-set of Rp categories populated the Rp− condition, and the other half of the items populated the Rp+ condition. Items in the Nrp categories were also randomly split between the Nrp_a and Nrp_b conditions.

Figure 3.

Behavioral paradigm. Initially, participants studied each one of the animal–name pairs in isolation once, with a parity judgment baseline task separating each presentation. Participants were then given the opportunity to retrieve the name of each exemplar out loud (yellow prompt) and restudy the correct pairing (blue prompt) twice in each of the following 4 rounds of interleaved retrieval practice and restudy. We manipulated the order of these 2 constituent tasks, such that Rp+ items were subjected to Competitive retrieval practice attempts followed by feedback (i.e., the yellow preceded the blue prompt, as in the case of Abner). Items in the Rp− and Nrp_a/b conditions (e.g., Odin), in contrast, underwent retrieval practice only after first receiving the correct answer in blue, making the retrieval task relatively Non-Competitive. A final opportunity to study all of the intact pairings once, without retrieval, was provided in the same manner as the initial study period, before the final cued-recall test. Note that Rp− and Nrp items were seen the exact same number of times in exactly the same fashion; the only difference between the conditions was that participants performed Competitive retrieval practice on other items from the Rp categories (but not from the Nrp categories).

Figure 4.

Derivation of the neural learning score. For simplicity, in this diagram, we only consider 1 animal category per Nrp/Rp condition, each with 4 exemplars (the outline colors indicate the sub-condition within that category, with pink representing Rp− and Nrp_a items and green representing Rp+ and Nrp_b items). For our primary analyses, we computed similarity matrices separately for items within each Rp category (Rp− to Rp+ similarity) and Nrp category (Nrp_a to Nrp_b similarity) based on data collected during the initial study period (S1). The within-category item similarity was expected to be comparable, on average, across categories prior to our behavioral intervention. Thus, items from S1 are represented as being equidistant from each other in the within-category similarity structures. In the figure, differentiation is shown as reduced similarity for items in the Rp category during the S2 phase, relative to the level of similarity that was present during S1. Our “neural learning score” summarizes these effects using a single number that reflects the change in similarity (from S1 to S2) for Rp items, relative to Nrp items. Because the values entering into the score are similarity (rather than distance) measures, we interpreted larger, positive scores as reflecting greater differentiation of the Rp items compared with Nrp items over time. Of primary interest was the relationship, across participants, between the magnitude of the neural learning score and the size of the revRIF effect, measured on the final recall test.

Figure 5.

Behavioral results. The left panel depicts Competitive retrieval-practice success for Rp+ items across the 4 intervening rounds between the initial study phase and the final restudy opportunity. While participants initially struggled to name the Rp+ animals in the Competitive retrieval condition, they managed to do so with greater success on subsequent retrieval practice attempts. The right panel depicts the final recall accuracy for all 3 conditions. Competitive retrieval practice facilitated Rp+ items above the Nrp baseline. There was a marginally significant trend for the Rp− competitors to be facilitated above the Nrp baseline, as well, indicating numeric revRIF across participants. Error bars represent SE of the mean across participants. *P < 0.05, †P = 0.05.

Figure 6.

Individual differences analysis linking brain and behavior. We found a significant positive correlation between our neural learning score derived from a left hippocampal ROI (in red) and the degree to which the Rp− items were facilitated above baseline on the final test. A similar trend was found when considering a bilateral hippocampal ROI (in blue). The right hippocampal ROI (yellow) showed a far less reliable trend in the same direction. *P < 0.05, ∼P = 0.10.

Figure 7.

Within-condition differentiation and revRIF. By examining the relationships between revRIF and the constituent parts of the neural learning score derived from the left hippocampal ROI (top row, red) and the bilateral hippocampal ROI (bottom row, blue), we observed reliable correlations between revRIF and the neural differentiation of Rp categories, but not for Nrp categories. *P < 0.05, **P < 0.01.