Competition between Associations in Memory

Knowledge often demands that we remember associations that share an item. Suppose you are learning about animals. You first find out that chickadees eat seeds, an association, AB, between seeds (A) and chickadees (B). Later, you find out that chickadees (B), in turn, are eaten by hawks (C), the association BC (boldface is used here to highlight the shared item). In a second example, you are keeping track of which teams have played each other in a children's sports tournament, to be able to plan future games. Team A (“Trojans”) played against Team B (“United”) who played against Team C (“Strikers”) in a later game. Again, this entails remembering both the AB and BC associations; there is no sense in which the later association replaces the earlier association. However, the repeated item (B) in two pairs introduces associative interference. Without mechanisms to specifically address this interference, mathematical models predict a somewhat mutually exclusive relationship between the AB and BC memories (e.g., Caplan, Rehani, & Andrews, 2014). That is, if AB is remembered, BC is less likely to be remembered and vice versa. This formal argument seems different than our daily experience. Clearly, we can remember new information related to B without losing the previous memories. Our overarching question is whether this competition ever materializes in human memory tasks and is then resolved, or possibly is never a challenge to begin with. Note that in these examples, the order of the two associations relative to each other might or might not be known. Here, we are focused on whether both associations can be remembered, or just one at the expense of the other, whether or not their relative order is also known.

Most associative interference studies have been modeled not on the AB/BC arrangement but on AB/AC learning, where the left-hand item is always the shared item (note that AB/AC is also an associative interference paradigm, but the shared item is always the cue, whereas in AB/BC, the cue switches positions). Moreover, most of the theory of associative interference has been developed with two-list procedures, where List 1 contains unambiguous pairs (all the A_iB_i plus control pairs) and List 2 introduces interference (all the A_iC_i plus more control pairs). One can look for evidence of competition by comparing accuracy, on average, of memory for interference pairs compared with control pairs with no repeated items. However, it has long been pointed out that there can be a general effect, where, for example, the interference pairs as a set are remembered worse than the control pairs as a set (e.g., Kliegl & Bäuml, 2021; Martin, 1971b; Postman, Stark, & Fraser, 1968; Underwood & Schulz, 1960). With two lists, this could be as simple as participants inhibiting all the response items of interference pairs in List 2 (or similarly in List 1), regardless of their specific pairings. Returning to our situation, where the overlapping item switches position (A_iB_i and B_iC_i), if interference occurs directly between pairs sharing an item, what is needed is a test of whether a particular A_iB_i pair competes with a particular B_iC_i pair—in other words, a correlation across overlapping pairs indexed by i. If A_iB_i and B_iC_i compete, then if one pair is remembered, it will often be at the expense of the other, producing a negative correlation between the pairs. If there is no competition at the level of pairs, we might instead find zero correlation, indicating independence (accuracy of B_iC_i is unrelated to accuracy of A_iB_i), or a positive correlation, indicating facilitation (A_iB_i and B_iC_i tend to both be remembered or both be forgotten). Empirical results actually show correlations of zero, indicating no competition but independence of the memories, or even positive correlations, indicating a facilitatory relationship between the memories (Burton, Lek, & Caplan, 2017; Tulving & Watkins, 1974; Delprato, 1972; Greeno, James, & DaPolito, 1971; Martin, 1971a, 1971b; Wichawut & Martin, 1971). This could mean that contrary to the intuition of many researchers, as well as memory models with competitive retrieval, pairs sharing an item never, in fact, compete in memory. Alternatively, some characteristics of the tasks may have enabled participants to overcome and sometimes even reverse competition between overlapping memories by the time memory is tested. In support of the latter, Caplan et al. (2014) were able to show unambiguous evidence for the presence of competition between overlapping pairs. The key was to construct lists of the form AB, BC, CD, DE, EF, FA (shuffled). In such “double-function” lists (Primoff, 1938), each item is a left-hand item of one pair and a right-hand item of another pair. When ambiguous pairs are segregated to different lists, as is frequently done with AB/AC paradigms, participants can use list-membership to resolve interference (Kliegl & Bäuml, 2021). However, with the double-function paradigm, which we adopt here (Figure 1), list-discrimination cannot be used because interfering pairs are studied within a single list. Pairs were tested with a two-response procedure.¹ For example, given B as a cue, participants attempt to recall A and C. In response to B as a cue, accuracy for BC was negatively correlated with accuracy for AB (“same-probe” correlation, where accuracies of AB and or BC were derived from the two responses to a single (B) cue; see Figure 1B, C). Follow-up data-analyses showed that it was not just the two responses (A and C) competing to be retrieved, but the memories for the AB and BC where associations were stored with somewhat mutually competitive strengths (the distinct-probe correlation described in the methods and elaborated in Caplan et al., 2014).

If the range of interference resolution is book-ended by this competition effect and classic independence with AB/AC learning, this suggests that associative competition is, indeed, present initially (evidence: negative correlation with the one-list double-function task), but is often resolved by the time researchers test memory (evidence: independence with the two-list AB/AC task). The single list in the double-function paradigm prevents participants from using list-membership to protect against associative ambiguity. In addition, the long chains of double-function pairs might explain why this task produces pair-specific competition whereas double-function lists with shorter, three-item chains has produced pair-to-pair independence or even net facilitation (Horner, Bisby, Bush, Lin, & Burgess, 2015; Horner & Burgess, 2013, 2014). Because interference was present but not complete, this paradigm is well positioned to investigate neural processes underlying both the materialization and resolution of competition between associations.

Previous studies of neurocognitive mechanisms of associative interference, mostly using AB/AC tasks, have reported interference on average but not tested for pair-to-pair competition. Those that have tested for pair-level effects have produced results consistent with the behavioral literature, confirming near-zero or positive correlations in pair-specific analyses (e.g., in AB/AC lists, Kuhl, Shah, DuBrow, & Wagner, 2010, and in short, three/four-element double-function lists, Horner et al., 2015). These studies have therefore focused on understanding how the brain produces good memories of both pairs, but a clear view of processes by which competition initially materializes has remained elusive. Here, we investigate the initial piece of the story when competition first emerges, by studying brain activity in a one-list double-function associative interference paradigm that does show evidence of competition at the level of pairs. In addition, whereas most neuroimaging studies of associative interference have recorded only during retrieval and encoding of the later-studied pairs, we also recorded and analyzed activity during encoding of the earlier-studied pairs. Even in our paradigm, interference is resolved for a substantial proportion of pairs. This might occur during study of the later pair. However, a hypothesis that has not yet been tested is that processes, already during encoding of the earlier pair, might make it more likely that later interference can be resolved.

We scanned participants while they studied double-function lists, tested with two-response cued recall (Figure 1). We sought activity related to interference and its resolution during study of the later pair, but also prospectively, during study of the earlier pair linked by a common item (A_iB_i and B_iC_i; due to randomization, the earlier pair could be either A_iB_i or B_iC_i). We tested the following nonmutually exclusive hypotheses, which build on and connect with prior studies that have traced what are presumably the later interference-resolution stages (reviewed in more detail in the Discussion section). These analyses specifically take advantage of the fact that we recorded brain activity during the earlier- and later-studied pairs.

Regarding the source of interference, our first hypothesis was that neurocognitive processes that lead to good memory also lead the pair to compete, which falls out of models that assume retrieval is competitive. Thus, if encoding strength is indicated by brain activity that shows a subsequent-memory effect (greater activity during later-remembered vs. later-forgotten pairs; Kim, 2011), that same activity should also be associated with competition between two pairs. In other words, the neurocognitive processes that lead to good memory also lead the pair to compete. Next, consider that associative interference studies have found evidence that participants remember the previous pair when presented with a pair with a repeated item (e.g., Richter, Chanales, & Kuhl, 2016; Horner et al., 2015; Kuhl, Rissman, Chun, & Wagner, 2011; Kuhl et al., 2010). Extending these findings to explain pair-specific competition in our paradigm, our second hypothesis was that retrieval during study produces interference, consistent with behavioral effects found by Caplan et al. (2014). Thus, retrieval-related activity during the later pair, such as is found in the precuneus in related paradigms (e.g., Himmer, Schönauer, Heib, Schabus, & Gais, 2019; Brodt et al., 2016; Wimber et al., 2008; Phillips & Niki, 2002), may be associated with mutually exclusive memory—one pair remembered at the expense of the other. Moreover, activity during the earlier pair that reflects this propensity to reactivate may then produce proactive interference. The two hypotheses are not mutually exclusive and might coexist. As elaborated in detail in the Results section, the former would be supported if regions showing the simple subsequent-memory effect contrast also appear in the interference contrast during the earlier-studied pair. Otherwise, the hypothesis will not be supported, but also not strictly rejected, because there could be activity beyond the sensitivity of our measure. The latter will be supported if regions that appear in the contrast aimed to isolate reactivation also appear in the interference contrast during the later-studied pair.

Regarding the resolution of interference, our first hypothesis was inspired a different body of work, associative inference. This paradigm has similar task design but very different research goals. In studies of associative inference, having studied AB and BC, participants must infer the association AC. Given the major role ascribed to the hippocampus in transitive and associative inference (e.g., Zeithamova & Bowman, 2020; Zeithamova, Dominick, & Preston, 2012; Bunsey & Eichenbaum, 1996), our first hypothesis was that activity in the hippocampus overcomes competition between associations and can possibly even reverse it. Our second hypothesis, which could coexist with the first, was that interference may be resolved when the participant thinks about (i.e., retrieves) the earlier-studied pair while viewing the later-studied pair. Simply retrieving the earlier pair, as mentioned in the previous paragraph, would be expected to exacerbate a negative correlation between encoding strengths of the two pairs, so for interference to be resolved, additional processes should be present. Previous studies have implicated ventromedial prefrontal cortex (vmPFC) in associative inference (e.g., Zeithamova & Bowman, 2020; Spalding et al., 2018; Zeithamova et al., 2012; Kumaran, Summerfield, Hassabis, & Maguire, 2009). A third hypothesis is possible and testable with data recorded during the earlier-studied pairs. That is, there might be cognitive processes during the earlier-studied pair that make its representation in memory conducive to resolution with the later-studied pair, reminiscent of prior-knowledge effects (Sommer, Hennies, Lewis, & Alink, 2022; Sommer, 2017). These three hypothesis are not mutually exclusive and might all coexist. They are each tested in their own right. The first interference-resolution hypothesis would be supported if hippocampal regions were isolated in the resolution contrasts during the later-studied pair. The second hypothesis would be supported if the same region or set of regions were found to be significant in both the reactivation and resolution contrasts during the later-studied pair. The third hypothesis would be supported if a region or regions showed robust effects in the resolution contrast during the earlier-studied pair.

Finally, once we obtained activity consistent with reactivation of the earlier pair while studying the later pair, we sought convergent evidence that reactivation was, in fact, occurring, using representational similarity analysis (RSA; Kriegeskorte, Mur, & Bandettini, 2008) and single-voxel correlations across trials. We also interrogated the nature of that reactivated activity, testing the hypothesis that later reactivation of memory of a pair reactivates different, higher-order, activity (nonoverlapping areas; Favila, Lee, & Kuhl, 2020) than the original on-line processing of the stimuli.

Thirty (20 women, 10 men, age 28.8 ± 3.5 years; target sample size was set in advance based on related studies with similar expected sensitivity; Caplan & Madan, 2016) healthy participants were recruited from the university community. Participants had normal or corrected-to-normal vision and reported no past or present psychiatric or neurological disorders. The study was approved by the local ethics committee, Board of Physicians, Hamburg, Germany. All participants gave written informed consent and received monetary reimbursement (10 €/h).

We first adapted the verbal paradigm used by Caplan et al. (2014) to pictures, similar to our previous studies on emotional associates (Fujiwara, Madan, Caplan, & Sommer, 2021; Caplan, Sommer, Madan, & Fujiwara, 2019; Madan, Fujiwara, Caplana, & Sommer, 2017). The task is illustrated in Figure 1. To increase power for the fMRI analyses, we omitted the single-function (control) pairs that were in the original design. There were 12 double-function pairs per list, an interpair active-baseline task, only one test per item, vocal responses instead of typed, and changes to the timing. The experiment was implemented with home-grown MATLAB code and the PsychToolbox (Kleiner et al., 2007; Brainard, 1997; Pelli, 1997) and CogToolbox (Fraundorf et al., 2014) libraries. The testing session began with practice outside the scanner (not analyzed), to familiarize the participant with the procedures. The experimenter ensured that the participants understood their tasks and were able to recognize the practice-list stimuli. Most participants then did 18 runs (full procedure relevant to a given study set; one with only 17 runs and two with only 16 runs due to failure to start the scanner, and one with 14 runs due to withdrawing early) with scanning during the study phase only.

Stimuli were nameable, colored line-drawing object images from Rossion and Pourtois (2004), with some stimuli removed by the authors when they were thought to be difficult for German participants to identify. A stimulus was never used on more than one list (including the practice list), and a fresh full random assignment of stimuli to lists was done for each participant. In each study set, 12 objects were randomly assigned to a set of 12 pairs, with the restriction that they comprised a ring structure (AB, BC, CD, DE, EF, FG, GH, HI, IJ, JK, KL, LA) wherein every word was the left-hand member of one pair and the right-hand member of another pair (related to the stimulus structure of Horner et al., 2015, closed-loop triads, but differing in that here, the “loop” is longer and all items are of the same material, objects). The classic finding of associative symmetry of cued recall of pairs explains some important aspects of our task design. That is, in a standard list of nonoverlapping pairs, when the left-hand and right-hand items are treated the same (Horowitz, Norman, & Day, 1966), forward (given A, recall B) and backward (given B, recall A) cued-recall accuracies are equal (Asch, 1969). Moreover, in past experiments, if each pair is tested twice, cued recall of a given pair in the forward and backward direction nearly always produces the same accuracy; in other words, there is a very high correlation between forward and backward cued recall, computed across pairs (Kahana, 2002; Rizzuto & Kahana, 2000, 2001). Both the equivalence of mean accuracy and high correlation were also confirmed with the double-function list-structure we use here (Caplan et al., 2014).

Pairs were presented sequentially in random order. The items of each pair were displayed simultaneously, with the two items separated by a space in the center of the screen. Each pair was displayed for 3960 msec (2 × repetition time [TR] = 1980 msec), followed by a 150-msec blank interpair interval. Following the blank ISI, participants completed an active baseline task (described below), lasting from 2- to 4-integer multiples of 2/3 TR (step size 1.32 sec; range: 2640–5280 msec) to introduce randomly selected jitter into the timing. Note that the onset of the picture was not always at the same time with respect to the scanner pulse.

Scanning halted during the test phase so that vocal responses could be recorded. Each item served as a cue exactly once, requesting up to two responses, presented in random order. Each test trial, preceded by a block of the active baseline arrow task as in the study phase, consisted of a cue word centered on the screen. The phrase “Bild #1” (translation: “Picture #1”) was displayed centered, underneath the cue object while the participant was asked to vocally recall a word or phrase describing one of the two images that were targets of the cue object. The vocal response was recorded for 7850 msec into a sound file, but also scored in real time by the experimenter and a research assistant on a scoring sheet printed out in advance of the session. The recording was voice-activated, and the onset time was logged as well. A second response was collected the same way, with “Bild #2” displayed. Following previous implementations of the two-response procedure, participants were told they could give the two responses in any order they chose (Caplan et al., 2014; Barnes & Underwood, 1959). Accuracy was determined by matching the response with stimuli in the word pool. A response was considered correct if it was one of the two responses given, regardless of the other response and regardless of whether it was the first or second response given to the cue.

We were able to confirm that our task produced associative interference that could not be entirely resolved by the time memory was tested, replicating the central findings of Caplan et al. (2014) with the changes described above (see Results section).

To suppress rehearsal and reduce rest-related hippocampal activity (Stark & Squire, 2001), participants viewed an arrow pointing left or right, and responded with the button box with the button congruent with the arrow direction. Each arrow-task trial lasted a fixed duration (2/3 TR), and the number of trials was selected to fill the interpair jitter interval.

Correlations between pairs of accuracy outcomes were evaluated with Yule's 𝒬, equivalent to a gamma correlation for bivariate data (Kahana, 2002), but can otherwise be interpreted much like Pearson correlation; 𝒬 = 0 indicates statistical independence, 𝒬 > 0, positive coupling between the variables, and 𝒬 < 0, negative coupling, or some level of mutual exclusion (one memory tends to be recalled at the expense of the other). Statistics were conducted on log-odds-transformed 𝒬 values (logits), for which residuals are theoretically approximately Gaussian, thus appropriate for parametric tests, and resulting p values are the same as if one conducted a χ² test on the same contingency table (Hayman & Tulving, 1989; Bishop, Fienberg, & Holland, 1975).

Yule's 𝒬 is computed from 2 × 2 contingency table composed of tallies. As illustrated in Figure 1C, accuracy of one pair is in rows and accuracy of the other pair is in columns. If we label the four cells such that Cell A counts the number of trials for which both pairs are correct, Cell B when the first pair is correct and the second is incorrect, Cell C when the first pair is incorrect but the second is correct, and Cell D when both pairs are incorrect, 𝒬 = (ad − bc)/(ad + bc). The main relationship we are interested in is 𝒬_same−probe, where accuracy of two pairs sharing an item is derived from the test trial where the common item was the cue (for pairs AB and BC, we would use accuracy from the trial for which B was the cue, and both A and C were requested as responses). An example where the earlier pair was evaluated in what would be equivalent to a forward probe in a noninterference list of pairs (given B, did the participant recall C?) and the later pair was evaluated in a backward probe (given B, did the participant recall A?) is illustrated in Figure 1C. If associative interference is present, 𝒬_same−probe would be expected to be negative. If the interference between the earlier and later pairs is reversed, we would expect 𝒬_same−probe to be positive. Following Caplan et al. (2014), we also compute 𝒬_{distinct−probe}, where accuracy of the earlier pair is derived from a different test-cue trial than accuracy of the later pair. This exploits the fact that each item was given as a cue one time, with spaces for two valid responses. Consequently, each pair is tested twice, once in the forward direction, on the trial where its left-hand item is the cue, and once in the backward direction, on a different trial where its right-hand item is the cue (Figure 2A). Thus, the distinct-probe correlation is computed from a contingency table assembled from the relationship between pairs sharing an item, where accuracy of one pair was evaluated on a different test trial than accuracy of the other pair (Figure 2B). For example, we can assess memory for AB on the trial with A as the cue (correct if B was one of the two responses given and incorrect otherwise, corresponding to the forward response (Figure 2A). This would be yoked to the assessment of memory for BC on the trial with B as the cue (correct if C was one of the two responses given and incorrect otherwise; Figure 2A). In this example, both pairs were tested in the forward direction (on different trials). This pair of pairs would thus increment, in the contingency table, Cell A if both were correct, Cell D if both were incorrect, and Cells B and C if one were correct and the other incorrect (Figure 2B). The other relevant cases are where both pairs were tested in the backward direction (on different trials). Thus, 𝒬_{distinct−probe} also measures competition between memories of two associations sharing an item, but it eliminates the contribution of immediate competition between two candidate responses to a single cue, because memory accuracy is evaluated based on two different test-cue trials. Finally, we compute a control correlation, 𝒬_control, which is a bootstrap, computed on a contingency table composed of pairs from the same list that do not share an item (e.g., AB and CD). This estimates the positive correlation expected due to variability across lists (Caplan et al., 2014; Hintzman, 1980). If response candidates compete in response to a single cue, 𝒬_same−probe would be more negative than 𝒬_control. If, in addition, memories of two pairs sharing an item have been encoded in a competitive relationship, as found by Caplan et al. (2014), 𝒬_{distinct−probe} will also be more negative than 𝒬_control; otherwise, 𝒬_{distinct−probe} would be equivalent to 𝒬_control.

fMRI was performed on a 3 T system (Siemens Trio) with a 32-channel head coil. An echo planar imaging T2*-sensitive sequence in 64 contiguous axial slices (2 × 2 × 2 mm); TR, 1.98 sec; echo time, 26 msec; Multiband 2; parallel acquisition techniques factor 2; flip angle, 70°; matrix 64 × 64) was employed. High resolution (1 × 1 × 1 mm voxel size) T1-weighted structural MRIs were acquired for each subject using a 3D magnetization prepared rapid gradient echo sequence. Functional imaging data were processed using the SPM12 software (Wellcome Department of Cognitive Neurology; https://www.fil.ion.ucl.ac.uk/spm). Functional images were realigned and unwarped to correct for susceptibility-by-movement artifacts. For quality control, it was then checked whether individual participants had excessively moved within run and the normalization was checked via comparison of the template and normalized T1 using the contour-function in SPM. The anatomical images were coregistered to the mean functional image of that participant. The anatomical images were then segmented and transformed into standard stereotaxic Montreal Neurological Institute (MNI) space using Diffeomorphic Anatomical Registration Through Exponentiated Lie Algebra (DARTEL) as implemented in SPM12 and the deformation field applied to the functional images of the same participant. Functional images were smoothed with FWHM of 6 mm.

Individual subjects and group level data were analyses using the general linear model as implemented in SPM12 in a mass univariate approach. Here, we describe the full first-level model, with eight regressors of interest. In the Results section, we describe, in turn, each second-level model that is derived from those eight regressors.

For the first-level model, we sorted the encoding trials according to the subsequent performance in the same probe forward test, that is, current and competing pair remembered, only current but not competing pair, not the current but only the competing, and neither of the pairs remembered (Figure 1C). In other words, we sorted trials according to their Yule's 𝒬 cell in the same-probe forward/backward relationship (Figure 1C). We focused on the forward/backward relationship (i.e., the cue that tests the current pair in the forward direction and the competing pair in the backward direction) because of the expected high associative symmetry (Kahana, 2002), which was previously confirmed for this paradigm using verbal stimuli (Caplan et al., 2014) and—foreshadowing our current results—also observed in this experiment (Figure 3B). The symmetry characteristic ensures that the results of one test direction generalize to the other. Although the same-probe and distinct-probe correlations were significantly different from one another (Figure 4), we focused on the same-probe relationship for two reasons. First, the number of trials differentiating the same- and distinct-probe pairings was too small to be able to reliably identify brain activity that might be unique to the same-probe activity. Second, when we reran the analyses based on the distinct-probe relationship, activation maps were quite similar to those based on the same-probe relationship. We present only the same-probe results, as they should include both competition both at the item level (competition between the two response candidates to the cue item) and at the association level (competition between memories of two associations sharing an item).

Moreover, we dissociated “earlier-studied” pairs from “later-studied” pairs. “Earlier-studied” pairs were studied before the competing pair had been studied (AB, in the case of A as a probe in the example illustrated in Figure 1C). “Later-studied” were studied after the competing pair had been studied (LA, in the case of A as a probe).² Based on these considerations, the first-level model included eight regressors (earlier- vs- later-studied pair × 4 Yule's 𝒬 cells) that were created by convolving the onsets of the conditions with the canonical hemodynamic response function. The Yule's 𝒬 cells were the four cells of the contingency table used to compute 𝒬 (see Figure 1C). In addition, six movement regressors were added as nuisance variables. The encoding runs of each participants were concatenated with appropriate adjustments to the runs-specific constant, the autocorrelation structure, and the high-pass filter.

On the second level, we contrasted the parameter-estimates of the these regressors with participant as a random factor to identify brain areas where activity exhibited contrasts consistent with (re-)encoding, interference, and resolution during processing of earlier- and later-studied pairs. The nonsphericity correction for violation of the independent and identically distributed assumption was applied. The particular contrasts applied will be detailed in the Results section (see also Table 1). The fourth cell of the Yule's 𝒬 contingency table (both pairs incorrect) is ambiguous. Both pairs could be forgotten because neither was studied well, or because both were studied well but the two pairs competed such that neither response could be produced in response to the cue. For this reason, this fourth cell was usually left out of the contrasts, but beta values nonetheless plotted alongside the beta values for the other cells when illustrating the results.

It is important to note that the presence of competition, its neutralization, or reversal is evaluated by computing correlations across pairs. We cannot infer whether competition between memory of any one AB pair was resolved with memory of its corresponding BC pair. The way we have structured our analyses should be viewed in terms of working on the assumption that, for example, more cases for which both pairs are remembered reflect resolution of interference than cases for which one, but not the other, pair is remembered. No different than other contrast-based analyses of neuroimaging data, the results should be viewed with this limitation in mind. If two conditions in a contrast were truly sampling from a single distribution and differed only because of random noise, the contrast would usually be nonsignificant.

As a follow-up analysis, we conducted four psycho-physiological interactions (PPI; Friston et al., 1997) analyses using the results of the second-level resolution contrast, namely, left and right angular and mid frontal gyri (from Figure 6B) as seeds (thresholded p < .05, corrected) and compared coupling during resolution (Condition 1) with retro- and proactive interference (Conditions 2 and 3). Parameter estimates of the individual PPIs were tested on the second level using a one-sample t test. Because we did not a priori select which areas to use as seed regions, these PPIs should be viewed as exploratory.

Whereas the analyses just described identify changes in mean activity across conditions, the following set of analyses identify activity that varies across trials within condition. Our first question was whether we could find convergent evidence in support of the idea that memory of the earlier-studied pair is, in any concrete sense, retrieved (reactivated) during the later-studied pair, as other neuroimaging studies have found. Then, we asked whether later reactivation of memory of a pair reactivates the same or different activity (nonoverlapping areas) as on-line processing of the stimuli (Favila et al., 2020).

To get activity estimates for each individual trial as input for these analyses, detailed in the following sections, for each trial, we created an independent first-level model with one regressor containing only the corresponding trial, that is, its onset convolved with the canonical hemodynamic response function, and one for all other trials in that fMRI run (Mumford, Turner, Ashby, & Poldrack, 2012). Again, six movement regressors were added as nuisance variables; the correction for autocorrelation and a high-pass filter were applied. The t maps testing the beta of the trial of interest in each model against the implicit baseline was used for the following RSA to reduce the influence of noisy voxels (Dimsdale-Zucker & Ranganath, 2018). To maintain consistency with the other fMRI analyses, we used smoothed single-trial data with the same 6-mm FWHM kernel because it has been shown that that smoothing does not decrease the sensitivity of RSA (Hendriks, Daniels, Pegado, & Op de Beeck, 2017; Kriegeskorte, Cusack, & Bandettini, 2010; Op de Beeck, 2010).

We used RSA to evaluate voxel-pattern similarity between pairs of study trials. In contrast to the previous contrasts and the following single-voxel series of analyses, RSA analyses identify activity that produces a “voxel-pattern,” that is, carrying information in the relative weightings across voxels (e.g., Kriegeskorte et al., 2008; Haynes & Rees, 2005) that would traditionally be considered close enough together comprise a single “region.”

The first RSA was conducted to identify areas involved in on-line processing of the objects and the second to identify areas involved in successful reactivation from memory of earlier-studied pairs while studying the later pairs. In both RSAs, we employed a whole-brain searchlight approach (radius 5 voxels) and correlated (Pearson correlation across voxels) the resulting vectors of trial-specific t values across conditions of interest where only t values of trials within the same run were correlated. These correlation coefficients were averaged after Fisher z transform and were saved as value for the center voxel of the current searchlight.

With the first RSA, we aimed to identify areas involved in processing the objects based on perceptual similarity of the two trials sharing an object. Therefore, we correlated activity patterns of pairs with one overlapping item (e.g., BC with AB and CD in Figure 1A) and contrasted these against the correlation of each pair with all other pairs in that run (e.g., BC with DE, LA, EF, etc.; Figure 1A). The correlation coefficients of both conditions were contrasted in a paired t test as implemented in SPM12.

With the second RSA, we aimed to identify areas involved in successful reactivation of the competing pair. The rationale of this RSA was to identify areas where the similarity of activity patterns during encoding of later- and its competing earlier-studied pairs was greater when reactivation was presumed to have taken place versus no evidence of reactivation. These two pairs (e.g., BC and AB in Figure 1A) share the overlapping item (e.g., B), which will result in the same degree of similarity in all of the 4 Yule's 𝒬 cells. However, only when the competing pair is reactivated during processing the later-studied pair is there additional similarity expected in areas that are involved in processing and reactivation of these pairs (e.g., AB is reactivated during BC-studying resulting in similarity of brain activity with AB). The resulting correlation coefficients in each of the 4 Yule's 𝒬 conditions were contrasted on the second level using SPM12.

Pattern similarity can be caused not only by distributed patterns of activity but also by univariate activity differences between conditions because a stronger signal in a condition could lead to stronger correlations of their trials (Wagner, van Buuren, Bovy, & Fernández, 2016). To rule out this confound, we averaged the activity in spheres (radius 5 voxels) around the peaks identified by the RSAs. First, we contrasted this mean activity between conditions to test for differences in univariate activity that would not have survived the correction for multiple comparisons applied to the univariate analyses described above. Then, we correlated the individual difference in mean activity between conditions with the individual difference in pattern similarity.

Finally, we conducted single-voxel analogues of the two RSAs. This series of analyses tests for regional activity that might reflect perceptual processing and reactivation regardless of whether that activity produces a “voxel-pattern.” If a voxel reflects activity that is reactivated (whether high or low in value), even if there is no difference in mean activity between successful and unsuccessful reactivations, its activity may covary between pairs that are successfully reactivated. For each voxel individually, we correlated activity between the earlier and later pair sharing an item, across such pairings. We contrasted those correlations with correlations across pairings that did not share an item to identify voxels related to perceptual processing. We also contrasted correlations of voxel-activity between pairs sharing an item for which the earlier pair was later recalled versus not recalled, to assess possible reactivation of single-voxel activity. The larger the correlation, the more within-condition variability of activity between pairs of trials. This across-pair, within-condition correlation analysis was done on the single-voxel level but also using a searchlight approach (radius 5 voxels) and averaging the correlation coefficients within the spheres. The resulting correlation maps were contrasted between conditions on the group level as in the RSAs using SPM12.

Results of all fMRI analyses were considered significant at p < .05, FWE corrected for multiple comparisons across the entire scan volume or within the a priori defined anatomical ROIs. Based on the previous literature, ROIs for the univariate analyses were the hippocampus, precuneus, and vmPFC, for the mulivariate analyses in addition the inferior temporal and fusiform gyri. Bilateral hippocampus, bilateral precuneus, bilateral inferior temporal gyrus (all three subsections combined), and bilateral fusiform gyrus were computed from the Harvard-Oxford cortical and subcortical structural atlases. A vmPFC ROI was manually traced on the mean T1 image based on previously published postmortem data (Mackey & Petrides, 2014) using ITK-SNAP 3.6.0 (Yushkevich et al., 2006).

We first report behavioral results showing that the prior finding of pair-specific competition could be replicated with a pure double-function design (and thus, twice as many double-function pairs per list as in the mixed lists used by Caplan et al., 2014), with particular attention to whether associative interference is present, rather than being largely resolved. We also examined whether there was a predominance of proactive or retroactive interference in the behavioral data. Accuracy was in the middle of the allowable range (Figure 3), comfortably far from ceiling and floor, conducive to examining modulation of accuracy by competing pairs. There was little effect on accuracy of test position (Figure 3A) or direction (Figure 3A, B), replicating symmetry of mean cued-recall accuracy (Asch & Ebenholtz, 1962; Kahana, 2002). Next, consider that each item was used as a recall cue just once, but two responses were collected. This means that for a single pair, BC, the outcome of forward cued recall of BC can be evaluated by checking whether the participant produced C in response to B as the cue. Backward cued recall is evaluated on a different test trial: Given C as the cue, did the participant produce B as one of the responses? Tallied in this way, the correlation of forward and backward recall of individual pairs was high; Q = .87 (SEM interval: [−.017, +.015]), extending previously observed associative symmetry (e.g., Sommer, Schoell, & Büchel, 2008; Sommer, Rose, & Büchel, 2007; Caplan, 2005; Kahana, 2002; Rizzuto & Kahana, 2001) to pairs of pictures. More importantly, this shows that associative symmetry holds even in the presence of heightened competition (Caplan et al., 2014; Rehani & Caplan, 2011). Thus, we can safely collapse over test position and test direction in the remaining analyses. There were large effects of serial position on accuracy (Figure 3B). However, plotted differently, Figure 3C shows that the driving factor was not serial position, but rather the amount of interference present while the pair was studied. Pairs that were presented before either constituent item had been studied were most accurate, followed by pairs for which one, but not the other item had been seen, and the lowest accuracy for pairs for which both items had previously appeared in other pairs. In other words, proactive interference is a major source of variability in accuracy in this task. Figure 3D shows the breakdown of accuracy as a function of cell within the contingency table from which 𝒬_same−probe is computed (Figure 1C), that is, between the earlier-studied and later-studied of two pairs sharing an item. The first and fourth conditions are cases where both pairs are recalled or both pairs are not recalled, respectively. The middle conditions indicate competition between memory for the two pairs. Inspection of those middle bars shows that it was more common for the earlier-studied pair to be remembered at the expense of the later-studied pair, than the other way around. Consistent with Figure 3C, proactive interference was more common than retroactive interference.

Analyses of the interference-related Yule's 𝒬 values (Figure 4, explained in the Methods section and illustrated in Figures 1 and 2) confirmed the presence of direct competition between pairs sharing an item. Namely, both 𝒬_same−probe and 𝒬_{distinct−probe} were significantly below 𝒬_control, t(29) = −6.57, p < .0001, and t(29) = −5.28, p < 0.0001, respectively, replicating Caplan et al. (2014). 𝒬_same−probe was also significantly more negative than 𝒬_{distinct−probe}, t(29) = −3.25, p = .0029, a novel finding suggesting the presence of both simultaneous competition at time of test between two candidate target items, and a competitive relationship between memory for the pairs, themselves.

The conceptual replication of the negative correlation between overlapping pairs extends the boundary conditions for this result. Modeled on the task used by Caplan et al. (2014), our paradigm differed in several ways: 1) The stimuli were drawings of objects instead of words. 2) Recall was vocal rather than typed. 3) To maximize data-yield to support the analyses of interest, we omitted single-function pairs and doubled the number of double-function pairs. Despite all these changes, recall of pairs sharing an item was negatively correlated, indicating that as in Caplan et al. (2014), pairs competed directly in memory. The correlation was significantly negative (differing from behavioral findings from associative interference paradigms as described in the Introduction), but not as negative as possible (−1). This satisfied the initial conditions we sought: Associative interference was present, and partly, but not completely resolved by the time memory was tested.

As described in the Methods section, the first-level model included eight regressors of interest (earlier- vs. later-studied pair × 4 memory-outcome conditions corresponding to the quadrants of the Yule's 𝒬 table; Figure 1C). Isolating activity during earlier-studied pairs identifies activity during encoding that results in either proactive interference (only the current but not the later-studied pair will be remembered) or in resolution of interference of the current pair with the later-studied pair (both associations will be remembered). Activity during encoding of later-studied pairs (similar to AC in classic paradigms) can be related to either proactive interference of the earlier studied pair, retroactive interference of the current pair, or resolution with the earlier-studied pair. We first report analyses of activity during the earlier-studied pair and then during the later-studied pair.

The order of the regressors representing the 4 Yule's 𝒬 conditions was always (1) current and competing pair remembered, (2) only current pair remembered, (3) only competing pair remembered, (4) neither pair remembered. In the text, because our main analyses are restricted either to the earlier- or to the later-studied pair, we use shorthand, noting only the first four (earlier-studied) or last four (later-studied) regressors.

It should be borne in mind that the following contrasts are not designed to be mutually exclusive, but rather, to identify particular relationships of regional activity to the task. As we will note, when a region appears in one contrast, whether it does or does not appear in another one may further specify its putative role in the task.

Before delving into effects specifically related to interference and its resolution, we conducted a simple subsequent-memory effect analysis, to identify the “basic” encoding regions. We identified areas showing a general subsequent memory effect for the current (earlier-studied) pair irrespective of memory for the competing (later-studied) pair by the contrast [1 1 −1 −1]. A set of regions comprising bilateral anterior hippocampus, vmPFC, and visual areas showed robust subsequent memory effect (SME) (Table 1, “encoding”).

To identify areas where activity during encoding of the current pair resulted later in interference with the competing pair, that is, that showed greater activity when the current pair was remembered but the competing pair was not remembered, the contrast [−1 2 −1 0] was applied. The case of both pairs forgotten was omitted from this contrast because it is ambiguous whether such cases are due to a failure of a resolution attempt or that one or both associations were individually not remembered. Regions within posterior hippocampus on both sides showed this effect (Table 1, “interference,” and Figure 5A), suggesting that particular hippocampal-dependent study processes produce a memory that eventually will compete with encoding of the later-studied pair. The inferior frontal and lingual gyri also showed this effect.

Following up on these findings, the contrast [−1 1 0 0] specifically tests for greater activity when only the current, and not both pairs, will be remembered. Only a right posterior hippocampal region showed this effect. We did not observe any areas showing activity related to retroactive interference from the later-studied competing pair (contrast [−1 −1 2 0] as well as [0 −1 1 0]), consistent with the small corresponding behavioral effect (Figure 3D).

Next, we ask if there is any activity present during processing the earlier-studied pair that later results in resolution with the competing pair (Table 1, “resolution,” and Figure 5B), testing our third hypothesis about the cause of resolution. The contrast [2 −1 −1 0] isolates activity associated with both pairs being subsequently remembered, versus only one pair remembered but not the other. Again, hippocampal subregions bilaterally, most extensively in the anterior subdivision, showed greater activity during successful encoding of the earlier-studied pair when the later-studied competing pair was also successfully remembered. Numerous other regions, most importantly the vmPFC and ventral precuneus, showed a similar pattern of activity.

Following up on these findings, the contrast [1 −1 0 0] specifically tests for greater activity when both pairs were remembered versus the current pair remembered but the (later-studied) competing pair not. In other words, given that the current pair was remembered, was the competing pair also remembered or not? This contrast also isolates resolution-related brain activity during study of the earlier of the two pairs. Anterior hippocampus, vmPFC, and precuneus were also found in this contrast.

Paralleling the analyses of the earlier-studied pairs, we first tested, for the later-studied pairs, which brain areas show a simple SME (contrast [1 1 −1 −1]), which revealed a set of regions comprising the supramarginal, middle temporal, and fusiform gyri but not the hippocampus (Table 2, “encoding”). However, this might be simply an effect of lower power because there was less successful encoding of the later-studied pairs, in particular whereas the proportions of later-studied pairs in the first regressors (Yule's 𝒬 cell 1) was similar to the earlier-studied pairs (Figure 3D), there were substantially fewer in the second regressor due to proactive interference. Moreover, this contrast confounds interference effects with subsequent-memory effects, as we shall see below when we seek activity related to encoding or “re-encoding of earlier and encoding of later-studied pair.”

Next, we looked for regions that might reflect retrieval of the earlier-studied pair during study of the later pair (although this should be viewed not as conclusive; for convergent evidence, see the follow-up correlational analyses below). Areas that showed a reactivation-like pattern of activity were identified by the contrast [1 −2 1 0], on the logic that if the earlier pair was remembered while studying the later pair, it is more likely to be recalled correctly during the subsequent memory test than if the earlier pair were not remembered during the later pair. Moreover, reactivation offers more encoding time to the earlier pair, which should also increase the probability that the earlier pair would later be recalled. Regions showing this pattern included the vmPFC, precuneus, middle frontal, and angular gyri (Table 2, “reactivation”). Because these analyses measure the amount of activity increase rather than the information-content of that activity, this set of putative “reactivation” regions might reflect neural processes that evoke reactivation rather than the reactivated information, itself, which might be housed elsewhere. The RSA analyses will show some convergent support for this latter interpretation.

During study of a later pair, retrieval of the earlier pair could result either in interference with the current encoding or in resolution of interference, which we target in the next contrasts. Regions that appear in the current “reactivation” contrast and the “interference” contrast are candidate regions for introducing interference via reactivation. Likewise, regions that appear in the current “reactivation” contrast and the “resolution” contrast are candidate regions for resolving interference by acting on memory of the earlier pair.

We next asked if there is any activity present during processing the later-studied pair that reflects proactive interference from the earlier-studied pair (Table 2, “interference,” and Figure 6A). We applied the contrast [−1 −1 2 0] to identify areas where only the competing but not the current pair will be remembered, excluding the ambiguous case where both pairs are forgotten. In addition to the vmPFC and precuneus that also showed reactivation-like activity patterns, the bilateral ventral striatum showed greater activity when only the competing, earlier-studied but not the current pair was successfully encoded. Very similar precuneus regions also appeared in a follow-up contrast, [−1 0 1 0], although not reaching significance (p < .1), contrasting only the competing pair recalled versus both pairs recalled.

We pause here to emphasize how remarkable it is that precuneus, vmPFC, and ventral striatum regions showed such robust activity differences that were primarily due to memory for a pair studied at an entirely different time in the study phase. This is compellingly consistent (although not conclusive) with the idea that these regions retrieve prior memories, but if there is too much reactivation, this risks leaving little opportunity to resolve those retrieved memories with the current pair. Convergent evidence for the presence of reactivation is reported below, in the trial-to-trial correlational analyses. We did not observe any region where activity related to retroactive interference on the competing, earlier-studied pair (contrast [−1 2 −1 0]).

Next, we isolated areas where reactivation of the competing, earlier-studied pair resulted not in interference but possibly resolution with the currently processed pair (Table 2, “resolution” and Figure 6C). As with the earlier-studied pairs, the contrast [2 −1 −1 0] was used, contrasting both pairs recalled versus only one recalled. The precuneus/posterior cingulate as well as the middle frontal and angular gyri were active during probable resolution of interference. These regions also exhibited the simpler, follow-up contrast, [1 −1 0 0], both pairs recalled versus the current pair forgotten but the competing pair recalled. Our first resolution hypothesis, that the hippocampus produces resolved associations, was not supported (although it is possible that a hippocampal effect is present but underpowered). Rather, our second (although not mutually exclusive) resolution hypothesis was supported, namely, nonhippocampal regions more plausibly produce resolved associations.

Of the four PPIs, using the left and right angular and mid-frontal gyri as seed regions, only one reached significance individually (p = .014, Z = 4.30; but after a Bonferroni correction, was only a trend). This was characterized by the left mid-frontal gyrus seed region exhibiting stronger coupling with the precuneus during integration than interference ([−8 −50 36], arrow in Figure 6C).

The suspected presence of reactivation raises the possibility that the earlier, retrieved pair could be encoded during study of the later pair, or both could be encoded at that time. We wondered if activity in the hippocampus might reflect this associative encoding, where sometimes the current (later-studied) pair is encoded, other times the competing pair (earlier-studied, retrieved during the current trial) is instead being encoded, or both. We collapsed together trials for which either the current or competing pair was correct, and contrasted those with trials for which both pairs were forgotten (Table 2, “(re-)encoding,” and Figure 6C). In other words, the contrast [1 1 1 −3] expresses the idea that activity in a particular brain region might reflect the total amount of encoding occurring, whether it is devoted exclusively to the current pair, exclusively to the (reactivated) competing pair, or split somehow between the two. A region within the posterior hippocampus on each side had an activity profile very much like this, as well as a region within the middle temporal gyrus. Thus, posterior hippocampal and middle temporal activity during the later-studied pair could reflect encoding of whatever is in mind—the current or competing pair or potentially both. It is interesting that anterior regions of the hippocampus showed neither the “naïve” subsequent-memory effect, nor this (re-)encoding effect during the later-studied pair.

The set of regions that we identified with the “reactivation” contrasts is consistent with the idea that reactivation occurs, but the link to this interpretation is still quite indirect. If reactivation is, in fact, taking place, then as in previous studies, we should be able to observe some similarity in brain activity between the earlier- and later-studied pairs when reactivation presumably succeeded. The regions identified in the previous contrasts more likely reflect control processes that identify the repetition, initiate or retrieve the information, rather than the information-content of the reactivated memory, itself. In fact, it is also possible that those “reactivation” regions enhance later memory of the earlier-studied pair completely apart from any putative reactivation. Here, we ask if we can obtain more concrete evidence that the earlier pair is sometimes remembered during study of the later pair. Namely, is there similarity of brain activity between earlier- and later-studied pairs that might reflect the memory that is first constructed and then later remembered? (See, e.g., Koen & Rugg, 2016; Staresina, Alink, Kriegeskorte, & Henson, 2013; Lee, Kravitz, & Baker, 2011, for this approach). In addition, coupled to this: What is, in fact, reactivated? To tackle this question, we focused on a hypothesis from Favila et al. (2020) that reactivated information is different (nonoverlapping regions) than on-line perceptual processing of the stimuli. In other words, what the participant remembers is at a different level of representation, presumably higher-order, than visual-processing of the object-pairs.

We therefore conducted analyses seeking brain activity that reflected the hypothetical reactivation, itself (described in the Methods section). That is, we sought activity patterns, as well as individual voxels, that were common to the earlier and later pair sharing an item, when the earlier pair was subsequently remembered (presuming successful reactivation) versus not (less successful reactivation). We compared this to a control analysis that identified similarity of activity due to the common item being visually processed during both pairs.

The maxima of the perceptual similarity RSA were in the inferior occipital ([−42 −88 −2], Z = 5.82; [46 −80 −6], Z = 4.97) and fusiform gyri ([−34 −54 −16], Z = 4.80; [40 −54 −16], Z = 5.11) where the fusiform cluster were the most anterior (Figure 7A). This establishes the set of regions showing pattern similarity likely due to visual processing of the item that was common between the earlier- and later-studied pair.

Reactivation was tested as in the univariate analyses with the contrast vector [1 −2 1 0]. This revealed a cluster in left inferior temporal gyrus ([−52 −32 −22], Z = 4.42) and vmPFC ([−4 40 −24], Z = 3.96; Figure 7B). Thus, the reactivation regions were further in the ventral visual pathway and did not include those whose voxel-patterns reflected visual processing of the shared stimuli, similar to Favila et al. (2020).

To test whether the reactivation RSA results were driven by univariate differences effects, we first contrasted mean activity in the spheres around the peak voxels identified by the RSA between Conditions 1 and 3 with Condition 2. Second, we correlated the individual differences between these conditions with the corresponding differences in similarity (Wagner et al., 2016). (Note that this control analysis of mean activity within spheres is not sensible for the perceptual similarity RSA due to the fact that all trials are in both conditions lead to equal mean activity in both conditions.) In the inferior temporal gyrus, mean activity showed a trend toward a significant difference between conditions, t(30) = 1.55, p = .065, and this difference did not correlate with the individual difference in similarity between conditions (r = .094, p = .612). In the vmPFC, a similar pattern emerged, namely, a trend toward a significant difference between conditions, t(30) = 1.43, p = .082, and no correlation of the individual differences with similarity difference (r = .274, p = .136).

For the perceptual similarity analysis, the searchlight approach resulted in a very similar set of regions as the perceptual similarity RSA, that is, bilateral early visual cortex (calcarine [−14 −96 −8], Z = 5.52; inferior occipital gyrus [−50 −72 −2], Z = 5.72; [52 −72 −2], Z = 6.26) and fusiform gyrus ([−32 −48 −10], Z = 3.73; [24 −46 −12], Z = 4.56) and, in addition, the precuneus ([−6 −54 10], Z = 4.09). Without the searchlight, the single-voxel analysis produced less smooth results but were otherwise similar, also identifying the early visual cortex ([−12 −94 −8], Z = 6.09; [20 −88 −8], Z = 5.87) and the fusiform gyrus ([−28 −44 −12], Z = 4.10; [30 −40 −14], Z = 4.58).

For the reactivation analysis, the searchlight approach converged with the RSA analysis, revealing clusters in the left inferior temporal gyrus ([−54 −32 −24], Z = 3.35; and vmPFC ([−4 34 −26], Z = 4.34). Without the searchlight, the single-voxel analysis also identified the left inferior temporal gyrus ([−54 −38 −24], Z = 3.16) and vmPFC ([0 34 −20], Z = 3.4). However, those clusters were not significant after correcting for multiple comparisons.

Thus, both distributed patterns (RSA analyses) and regional activity (single-voxel correlations) reflect pair-specific activity that is present during study of the earlier pair and then reactivated during study of the corresponding later pair. The regions showing these effects are different and higher-order than those reflecting perceptual processing of the displayed stimulus.

With the first behavioral paradigm in which participants are challenged by pair-specific associative interference but cannot fully resolve it, we identified neural processes that explain how competition between associations materializes, and how it can be overcome. We review these main findings and discuss how they dovetail with findings from prior studies in which pair-specific interference has been largely resolved.

A rare feature of our paradigm, offering access to brain activity throughout the study phase, is that it allows us to follow proactive interference-related activity from the earlier-studied pair to the later-studied pair. Our first hypothesis, that simple competitive retrieval based on encoding strengths is the source of interference, was not supported; brain regions that led to good memory for the earlier-studied pair (“encoding” contrast in Table 1) did not show interference effects (“interference” contrasts). Consistent with the established importance of the hippocampus for associative memory (e.g., Caplan & Madan, 2016; Saksida & Bussey, 2010; Konkel & Cohen, 2009; Eichenbaum, Yonelinas, & Ranganath, 2007; Mayes, Montaldi, & Migo, 2007; Davachi, 2006; Rudy & O’Reilly, 2001; Cohen, Poldrack, & Eichenbaum, 1997; Nadel & Moscovitch, 1997; Rudy & Sutherland, 1989; O’Keefe & Nadel, 1978), hippocampal regions were associated with subsequent memory of the earlier-studied pair. However, different hippocampal subregions (more posterior) were related to future associative interference.

Instead, our second hypothesis was supported; the primary source of interference appeared to be retrieval of the earlier-studied pair while studying the later-studied pair. This echoes neuroimaging findings on pairs with repeated items. This includes associative interference paradigms for which pair-specific interference had been largely successfully resolved (e.g., Richter et al., 2016; Horner et al., 2015; Kuhl et al., 2011; Kuhl et al., 2010), as well as associative inference paradigms, where the participant's explicit goal is to combine information from both component pairs to answer inference questions (e.g., van Kesteren, Rignanese, Gianferrara, Krabbendam, & Meeter, 2020; Zeithamova & Bowman, 2020; Zeithamova et al., 2012; Zeithamova & Preston, 2010) and confirms behavioral evidence of such retrieval occurring in the one-list double-function paradigm (Caplan et al., 2014). Specifically, high activity in posterior hippocampus (Figure 5A), with insufficient activity in vmPFC, precuneus, and anterior hippocampus (Figure 5B), lay the basis for interference with the later-studied pair.

During later-pair processing, the “reactivation” and “interference” contrasts share regions with nearby peaks in a set of regions comprising vmPFC, precuneus, and striatum (Table 2, Figure 6). Precuneus and striatum have been implicated in memory retrieval (Clos, Schwarze, Gluth, Bunzeck, & Sommer, 2015; Schwarze, Bingel, Badre, & Sommer, 2013; Huijbers et al., 2012). While studying the later-presented pair, right precuneus, along with bilateral vmPFC, left angular gyrus and thalamus, and right middle frontal gyrus, was associated with good memory for the earlier-studied pair, regardless of memory of the current pair. This is remarkable, because the contrast completely omits brain activity during initial study of the earlier-studied pair.

Aligning these results with those from the interference contrast (later-studied pair), a left-sided vmPFC region and the right precuneus region recur in the interference contrast with nearby peaks. Although speculative, this pattern of findings reinforces our previous suggestion that strong encoding of the earlier-studied pair in posterior hippocampus together with weak encoding in the vmPFC-precuneus-anterior hippocampus might result in subsequent reactivation via the striatum and precuneus that reduces encoding of the later, competing pair (cf. Long & Kuhl, 2019).

As with the origin of interference, our paradigm also allows us to track processes supporting resolution of interference across time, from the earlier- to the later-studied pair. Our first hypothesis, extrapolating from transitive/associative inference (e.g., Horner et al., 2015; Zeithamova et al., 2012; Heckers, Zalesak, Weiss, Ditman, & Titone, 2004; Preston, Shrager, Dudukovic, & Gabrieli, 2004; Dusek & Eichenbaum, 1997; Bunsey & Eichenbaum, 1996), that the hippocampus supports successful resolution of the two pairs was unsupported. Perhaps the hippocampus was implicated in prior inference tasks for its role in supporting memory of the component associations (A → B and B → C), enabling a chained retrieval solution to the inference, or supporting encoding of the inferred A → C association during the B → C trial (Koster et al., 2018), without necessarily resolving any competition.

Our second hypothesis was supported. Although reactivation can produce interference, if there is not too much reactivation, the participant might sometimes use extra-hippocampal processes (Figure 6C) to resolve competition between the two co-activated associations, potentially producing facilitation between the memory for the two pairs (Burton et al., 2017; Wahlheim, Maddox, & Jacoby, 2014; Jacoby & Wahlheim, 2013; Wahlheim & Jacoby, 2013; Kuhl et al., 2010). Evidence for this can be seen in regions that appeared in both the “reactivation” and “resolution” contrasts (Table 2), which include a region within right precuneus, middle frontal, and angular gyri. Consistent with this, it is interesting that the activity level when both pairs were recalled (Figure 6A, R/R condition) was not the highest, but actually lower than when only the earlier (competing) pair was remembered and greater than when only the later (current) pair was remembered. Thus, if activity in the precuneus reflects the total amount of retrieval of the earlier pair while studying the later pair, this suggests that too much reactivation leaves too little opportunity for mid-frontal and angular gyrus to resolve interference. The angular gyrus has been linked to integration and retrieval of supramodal complex semantic knowledge (Gilboa & Marlatte, 2017) and multimodal feature integration during episodic retrieval (Bonnici, Richter, Yazar, & Simons, 2021). The computations of the angular gyrus, as part of a wider lateral parietal system, enable the on-line dynamic buffering of multisensory spatiotemporally extended representations (Humphreys, Ralph, & Simons, 2021; Xie, Li, Xie, Xu, & Peng, 2019). The mid-frontal gyrus has been implicated in attention (Bourgeois, Sterpenich, Iannotti, & Vuilleumier, 2022), which may be this region's specific role in resolving interference here.

Finally, our third hypothesis, which we were for the first time able to investigate, was supported. That is, activity during the earlier-studied pair apparently set the initial conditions for future resolution of interference. Among the regions that exhibited this effect (“resolution” contrast in Table 1) were vmPFC bilaterally, anterior hippocampus bilaterally, and precuneus. Similar regions were identified with the simple “encoding” contrast. This suggests that (contrary to the first interference hypothesis discussed in the previous section), a well-studied pair that is likely to be remembered is also one that has a good chance of being reconciled with an overlapping association in the future. This hippocampal activity was during the earlier-studied pair, so it does not reflect the resolution process, itself, but rather the formation of a memory with favorable properties for future resolution. The role of vmPFC here is consistent with numerous prior studies implicating vmPFC in encoding (Fujiwara et al., 2021) and in forming integrated representations (Gilboa & Marlatte, 2017). However again, vmPFC is not apparently carrying out the resolution, itself, but laying down the conditions for future resolution by other regions, namely, middle frontal and angular gyri and posterior cingulate, whereas the later competing pair is studied. Interestingly, an interplay of the vmPFC and hippocampus with the precuneus during encoding has been previously associated with incorporation of novel information into existing schemata (Sommer et al., 2022; Sommer, 2017). Liu, Grady, and Moscovitch (2017) found that an advantage for houses associated with famous faces versus nonfamous faces was attributable to activity in anterior hippocampus, vmPFC, and precuneus activity—similar to what we observed. The famous faces were presumed to provide richer representations to which the houses could be bound. In our task, this same set of regions may similarly provide additional details to the representation of the first pair, making it easier for the later pair to be reconciled with the earlier memory. This is in contrast to posterior hippocampus, which may produce a memorable association that is less amenable to resolution of competition with another pair. Importantly, participants do not simply resolve interference once it materializes, during the later-studied pair. Rather, at least as important, the way in which the earlier-studied pair is processed can be critical for subsequent successful resolution.

Having first identified the source of pair-specific interference and then identified brain activity related to the neutralization, or even reversal, of that interference, we now consider the cognitive processes that the latter activity might reflect.

First, we note that the term “integration” arises repeatedly in the associative interference neuroimaging literature, but with several meanings, each of which might be related to the neurocognitive processes we identified here. Integration can refer to the formation of a composite representation of two pairs in memory, such as encoding not just AB and BC, but something like ABC. This is the idea behind instructions to participants to form integrative imagery as a way to resolve interference (e.g., Burton et al., 2017; Anderson & Bell, 2001; Smith & Hunt, 2000; Anderson & McCulloch, 1999). However, such integrated representations in memory are hard, if not impossible, to confirm. Correct recall of both pairs (or even a positive correlation across pairs of pairs; both pairs remembered or both forgotten; Burton et al., 2017) could conceivably result from such an integrated representation, but this kind of result could have other plausible causes. Strictly speaking, the positive correlation only tells us that memory for A_iB_i and B_iC_i have a source of shared variance.

Consider retrieval of the earlier pair while studying the later pair. If the earlier pair is well encoded, it might be retrieved with little effort and rapidly, thus displacing very little encoding time from the later pair. In this way, a highly recallable earlier pair might facilitate encoding and subsequent recall of the later pair without any direct integration. Conversely, a poorly studied earlier pair, itself less likely to be recalled correctly, may require more study time to be recalled, thus also obstructing encoding of the later pair, making it likely that both pairs will not later be recalled. Our findings are somewhat in line with this; hippocampal activity that produced a subsequent-memory effect during the earlier pair was associated with good memory for both pairs.

Second, drawing an analogy to associative inference, resolution, and even reversal of interference in our task, when it does occur, might be caused by participants encoding the inferred association, AC, after retrieving the earlier pair while studying the later pair. This could positively couple the two pairs by adding a new retrieval route. Suppose that given B as a probe, BC were not remembered. If AB was remembered, A could then be used as a retrieval cue for C, via the encoded inferred AC association. Inferring the indirect association might be one role of mid-frontal and angular gyrus activity during the later pair. Then, posterior hippocampal activity, which appears to be an agnostic (re-)encoder during the later pair, could then store the indirect association if it were successfully produced, or else the retrieved earlier association or the current association.

A different notion is that ambiguity between similar stimuli may be addressed in part by changing the representations so that they are more orthogonal, termed “pattern separation” (e.g., Poppenk, Evensmoen, Moscovitch, & Nadel, 2013; Norman & O’Reilly, 2003; O’Reilly & McClelland, 1994; Marr, 1971). Interestingly, Becker (2017) found that the resolution of interference was solved by her model by making representations more similar rather than less similar (echoing the effects of strategy found by Burton et al., 2017), but the more general idea that representations may become more distinct (not necessarily orthogonal) when both associations are brought to mind together has been proposed by, for example, Kuhl et al. (2010, 2011), and Chanales, Dudukovic, Richter, and Kuhl (2019). Representational Hierarchical Theory, which assumes no special role for the hippocampus in memory, per se, implicates the hippocampus precisely in offering the brain the ability to discriminate stimuli that would be processed as highly similar by more upstream regions (Cowell, Barense, & Sadil, 2019; Bartko, Cowell, Winters, Bussey, & Saksida, 2010; Cowell, Bussey, & Saksida, 2010; Saksida & Bussey, 2010; Bussey & Saksida, 2002). This could neutralize interference by reducing the similarity-based ambiguity in encoded memories, but it might also reverse interference if the formation of distinctive representations were synergistic, likely to be successful for both pairs or unsuccessful for both pairs. The latter mechanism is, in fact, the antithesis of an integrated representation. The mid-frontal and angular gyrus activity during the later pair, related to success with both pairs, might contribute to distinctive encoding of one or both pairs, that is, however, unlikely to be like pattern separation.

If construction of distinctive representations is the main mechanism of resolution of interference, our findings during the earlier pair suggest that the distinctiveness process can begin even before both pairs are known. The anterior hippocampal and vmPFC activity during the earlier pair might already achieve some distinctiveness, enabling the participant to focus more on forming a distinctive representation of the later pair once it is presented.

This repertoire of possible mechanisms of resolution of interference suggests why our paradigm, in contrast to other associative interference paradigms, leaves some competition unresolved. Associative interference in other paradigms is typically in triad form, such as AB/AC. Constructing an integrated representation, ABC could result in AB and AC both being remembered (or both forgotten). Encoding the inferred, BC association, could result in good memory of both AB and AC. As already noted, assuming AC was not memorable, if the participant can retrieve B with A as a cue, the BC association offers a backup retrieval path to potentially produce C as well. In the three-item loops used by Horner et al. (2015), again, storing ABC or storing all component associations could both produce positive correlations in memory of AB, BC, and CA. With our larger ring structure, those approaches may resolve competition in one part of the ring, but at the same time increase competition in another part of the ring. Suppose the BCD is stored. That may positively correlate memory tests of BC and CD with one another, but it introduces an additional source of competition when testing the pairs AB and DE. Thus, both “integrative” solutions may explain why our participants can resolve some competition between overlapping pairs, but this reasoning also shows why it may be quite challenging to resolve all interference after only a single exposure to a list.

Forming more distinctive representations is a process that might have a benefit without such a cost. Increasing the distinctiveness of BC from CD will also be likely to increase the distinctiveness of BC from AB. In fact, as just suggested, this might be the mechanism by which activity during the earlier-studied pair increases the chance of resolution of interference, even before the later-studied pair is known.

A complementary set of analyses of trial-to-trial variability produced more direct evidence for the presence of reactivation of the memory of the earlier-studied pair while studying the later pair. These analyses also suggested the relevant reactivated information was at a relatively high level of representation, different from more face-value visual processing of the stimuli. To understand the logic here, assume that a region (either a set of individual voxels or a legitimate voxel-pattern) reflects high-order features of memorial representations. Those features might be high in value or low in value; hence, no difference is expected in average activity (and would be missed by the simple contrasts). What matters is whether the same value occurs both during the earlier pair and the later pair, when the earlier pair is brought back to mind. For example, suppose Region X reflects the amount of vividness of an image constructed to bind two objects together. One stimulus pair, AB, might be high in vividness; evidence of its reactivation would be high BOLD signal both during presentation of AB and during subsequent presentation of BC. A different pair, DE, might be remembered with verbal rather than imagery mediation. Consequently, BOLD signal in Region X would be low during presentation of DE. If X reflects information that is reactivated, that low-value BOLD level would be expected to reiterate itself during study of a later pair, EF. Thus, the prediction is not that Region X should (necessarily) exhibit greater activity when reactivation is successful versus not, but that its value should covary across trials (pairs) for which reactivation was likely to have been successful. This correlation should be greater than when computed across pairs for which reactivation was probably not successful. The results indeed supported similarity between the earlier-studied and later-studied pairs when reactivation would presumably have been more likely to have succeeded.

Moreover, the reactivation leading to subsequent memory of the earlier pair appeared at a high level of representation, further into the ventral visual pathway than on-line perceptual processing (Favila et al., 2020). Thus, participants do not, apparently, reactivate in the sense of re-imagining the two objects in the original pair with their detailed visual features, but rather, remember a highly processed, combined representation they had constructed of the two items. The control analyses confirmed that those were pair-specific distributed activity patterns (Koen & Rugg, 2016). Given the involvement of vmPFC, in particular, trial-to-trial variability in this region may reflect variability in producing the high-order representation of the association that is conducive to further elaboration. Such elaboration might, for example, support formation of an integrated representation or storage of the inferred, indirect association as just discussed.

Alternatively, if retrieval of the earlier-studied pair is decoupled from the low-level visual information in the stimulus, but rather, at a high level of representation, the more abstract representation might be more conducive to being transformed into a representation with more distinctiveness from other representations, or the high-level nature of the retrieval might indicate that the encoded representation of the earlier-studied pair may have often already been modified to be more distinctive.

To stick closely to the previous paradigm that produce unambiguous evidence of pair-specific competition (Caplan et al., 2014), we had to use verbal stimuli that would be conducive to recall. Our findings might be restricted to the verbal domain. To expand into nonverbal memory domains, it will first be necessary to adapt and validate the paradigm for forced-choice responses.

Our behavioral paradigm, with brain activity analyzed during the entire study phase, revealed that associative interference is not produced passively due to strength-based competition between overlapping memories, but rather due to proactive interference when the earlier-studied pair diverts encoding resources away from the later-studied pair, reflected in activity in the precuneus, among other regions. However, if the retrieval-related precuneus activity is not too strong, numerous additional regions, possibly coordinating with the precuneus, including angular gyrus and mid-frontal gyrus (but perhaps not hippocampus), can resolve interference. Finally, resolution of interference is enabled when the earlier-studied pair is studied in a particular way, involving activity in vmPFC and anterior hippocampus.

Reprint requests should be sent to Jeremy B. Caplan, Department of Psychology, Biological Sciences Building P217, University of Alberta, Edmonton, AB, T6G 2E9, Canada, or via e-mail: [email protected].

Data will be made available upon email request to the corresponding author.

Natural Sciences and Engineering Research Council of Canada (https://dx.doi.org/10.13039/501100000038).

Retrospective analysis of the citations in every article published in this journal from 2010 to 2021 reveals a persistent pattern of gender imbalance: Although the proportions of authorship teams (categorized by estimated gender identification of first author/last author) publishing in the Journal of Cognitive Neuroscience (JoCN) during this period were M(an)/M = .407, W(oman)/M = .32, M/W = .115, and W/W = .159, the comparable proportions for the articles that these authorship teams cited were M/M = .549, W/M = .257, M/W = .109, and W/W = .085 (Postle and Fulvio, JoCN, 34:1, pp. 1–3). Consequently, JoCN encourages all authors to consider gender balance explicitly when selecting which articles to cite and gives them the opportunity to report their article’s gender citation balance.