Repeated testing leads to improved long-term memory retention compared to repeated study, but the mechanism underlying this improvement remains controversial. In this work, we test the hypothesis that retrieval practice benefits subsequent recall by reducing competition from related memories. This hypothesis implies that the degree of reduction in competition between retrieval practice attempts should predict subsequent memory for practiced items. To test this prediction, we collected electroencephalography (EEG) data across two sessions. In the first session, participants practiced selectively retrieving exemplars from superordinate semantic categories (high competition), as well as retrieving the names of the superordinate categories from exemplars (low competition). In the second session, participants repeatedly studied and were tested on Swahili-English vocabulary. One week after session two, participants were again tested on the vocabulary. We trained a within-subject classifier on the data from session one to distinguish high and low competition states. We then used this classifier to measure the change in competition across multiple successful retrieval practice attempts in the second session. The degree to which competition decreased for a given vocabulary word predicted whether it was subsequently remembered in the third session. These results are consistent with the hypothesis that repeated testing improves retention by reducing competition.
How can learners maximize their retention of important information while minimizing time spent studying? Compared to repeatedly testing oneself, passively reviewing the to-be-learned material may seem like a relatively painless and effective way to confirm what is known and strengthen what is not. However, numerous studies have found that repeated testing leads to better long-term retention than repeated study1,2. In one notable experiment, Karpicke and Roediger3 trained participants to the point where they could recall items correctly, and then manipulated (as part of the same learning session) whether participants received additional memory tests or study trials for those items. When Karpicke and Roediger tested memory one week later, they found that delayed recall was much better in the additional-testing condition than the additional-study condition (80% recall after testing vs. 36% recall after restudy).
How can we explain this huge benefit of testing on delayed retention? A number of accounts for the testing effect have been offered in the literature, with varying degrees of support from neural and behavioral data1,2. One theory posits that the additional mental effort required for selective retrieval may reflect a desirable difficulty that enhances retention4,5. The close match between the cognitive processes involved during practice tests and the final test (i.e., transfer-appropriate processing) may also contribute to the benefit6. Another possibility is that searching for the correct response on a test leads to the elaboration of semantic associations through which the target material can later be accessed7. Yet another potential explanation is that repeated testing serves to reduce competition from related memories, thereby improving access to the target memory8.
Speaking to this last possibility, a large body of evidence highlights how selective retrieval (but not passive restudy) impairs the accessibility of competing memories. For example, repeatedly retrieving “SCOTCH” when cued with “ALCOHOL-S” impairs later recall of other alcoholic drinks (like “VODKA”) relative to unpracticed baseline categories. Simply reviewing “ALCOHOL-SCOTCH” facilitates recall of “SCOTCH,” but it is not associated with a corresponding impairment to the accessibility of “VODKA”. Since the first report of retrieval-induced forgetting (RIF)9, a number of explanations for this phenomenon have been posed10,11,12. Most commonly it has been attributed to the inhibition of competing memories9,13,14,15,16. Indeed, the results of a recent meta-analysis largely support the inhibitory account17, which notably has been implemented in a neural network model18 that has helped to explain and predict various neural and behavioral results19.
Should weakening of competitors be responsible for RIF, one might expect overall levels of competition to decrease over successive retrieval attempts; the relative reduction in competition should, in turn, predict competitor forgetting on a later test. Functional magnetic resonance imaging (fMRI) data support this prediction: One study found that, across multiple retrieval practices, decreased activity in anterior cingulate cortex (a region thought to reflect conflict between possible responses) predicted the degree to which participants showed an RIF effect; these particular changes in brain activity were selective to successful retrieval practice20. Another study found that, as unwanted intrusions from competing memories declined over the course of retrieval practice, so too did the extent to which their corresponding neural patterns were reactivated – all in a manner that predicted competitor forgetting21.
While the aforementioned results show that testing reduces competition, it has not yet been shown that this competition reduction supports the beneficial effects of testing on long-term retention of the tested items2. Here, we sought to test this hypothesis by obtaining a neural measure of competition, tracking its decline across retrieval practice attempts, and then relating this decline to subsequent memory for the practiced items (one week later).
To neurally measure competition, we leveraged prior work by Hanslmayr and colleagues that showed strong EEG differences for “high-competition” retrieval (in which participants selectively retrieved one of many exemplars from a superordinate semantic category) as compared to “low-competition” retrieval (in which participants retrieved the superordinate category given an exemplar)22. In our study, we trained a within-subject EEG pattern classifier on these two conditions to distinguish between high and low states of competition. This part of the experiment (Session 1) can be viewed as a kind of “competition localizer” that uncovers the neural signature of competition in EEG data.
After this initial “competition localizer” phase, we used a variant of the paradigm initially used by Karpicke and Roediger to show benefits of testing3: Participants were shown a set of Swahili-English word pairs and then were given multiple cycles of study and retrieval practice to learn these word pairs. In the original Karpicke and Roediger experiment, they compared several conditions that varied the amount of restudy and retrieval practice that occurred; in our study, our primary interest was in variance in competition within the retrieval practice condition, so we did not manipulate learning conditions across participants – we only included a single condition where participants received a fixed amount of retrieval practice and restudy, regardless of whether an item had been recalled successfully. We refer to this part of our experiment (Session 2) as the Swahili Learning Task. Note that competition in this task could arise from other studied items, or from knowledge acquired outside of the study context (e.g., the Swahili word “punda”, which means “donkey”, could activate the representation of “panda” because they sound alike); our approach is agnostic about which of these two sources of competition is affecting performance.
We recorded EEG during retrieval practice trials in Session 2. We applied the competition classifier (trained on Session 1) to these trials, thereby obtaining a trial-by-trial measure of competition. Lastly, we had participants return one week after Session 2 for another session (Session 3) where they were cued with Swahili words and asked to recall the associated English word. A schematic of the experiment with all three sessions can be found in Fig. 1.
For each word pair in Session 2, we computed the difference in the measured level of competition between the first successful retrieval practice trial for a given item and the last retrieval practice trial for that item; we call this measure “competition drop”. We chose the first successful retrieval practice trial for this measure because – like Karpicke and Roediger3 – we were interested in the learning processes that take place after the first successful recall attempt that determine whether or not the memory will be retained one week later. We hypothesized that the size of the competition drop value (i.e., the degree to which competition was reduced for this pair) would predict whether participants successfully recall that pair in Session 3. We tested this hypothesis on the pool of all practiced items, and additionally on subsets of items grouped by Session 2 retrieval accuracy, to test whether the competition drop metric held predictive power even when controlling for Session 2 accuracy. Our results support this hypothesis, as described below.
Average performance across participants is shown in Fig. 2a. On average, performance improved over each of the four retrieval practice trials (testing rounds) in Session 2 (R1–R4) but dropped between the last Session 2 testing round (R4) and the Session 3 test (T). The average performance on Session 3 was near 50%, which gives an (approximately) balanced number of remembered and forgotten items for analyzing which factors (in Session 2) predict subsequent recall in Session 3. Performance in Session 3 varied over subjects, as shown by the histogram of subject performance values in Fig. 2b. While the majority of subjects retained at least 50% of the Swahili-English word pairs, a substantial number retained less than that.
We can use the behavioral data from Session 2 to distinguish between subsequently remembered items (i.e., items correctly recalled in Session 3) and subsequently forgotten items. Figure 3a shows the average recall performance on each Session 2 testing round, grouped by performance on the Session 3 test. The blue and yellow bars show the average performance for subsequently forgotten and remembered items, respectively. Unsurprisingly, items that were better remembered in Session 3 were also better remembered in Session 2.
Next we examined how particular patterns of correct vs. incorrect recall (across trials) in Session 2 were related to recall in Session 3. For the purpose of this analysis, we encoded the Session 2 responses as 4-dimensional binary vectors, where a 1 in the ith position indicates that the participant responded correctly in the ith round of Session 2. As an example, a word for which the participant responded correctly in all but the first round would be encoded as 0111.
The result is shown in Fig. 3b. Some response configurations were much more likely than others; in general, once a participant was able to recall an item correctly, they continued to be able to do so in subsequent rounds. The more rounds in which the participant answered correctly, the more likely the participant was to remember that item in Session 3. Using a logistic regression classifier on the trials pooled over all subjects, we achieved 64% classification accuracy using recall accuracy for the four test rounds (in Session 2) as input features and subsequent recall accuracy (in Session 3) as the dependent measure. Chance is 50%. A permutation test with 1000 permutations yielded p < 0.001, demonstrating that Session 2 behavior was very significantly predictive of Session 3 recall.
Classifying High and Low Competition Neural States
We trained a classifier to distinguish between high and low competition trials from the Session 1 EEG data. Prior work using a similar paradigm found EEG signals that discriminated between high and low competition trials during the first 500 ms post-stimulus-onset22; however, we did not have a strong a priori hypothesis about exactly when (within this window) competitive states would be most distinguishable, and we did not want to rule out the possibility that slightly later time points would also discriminate. As such, we applied the classifier in a sliding window fashion, retraining and testing for each time point independently. The features given to the classifier were the voltages at each of the 64 EEG channels, averaged over a 50 ms window. We were able to successfully decode over a large period of time post-stimulus-onset, shown in Fig. 4. Significantly-above-chance decoding started at the 60 ms window (encompassing data from 60 ms to 110 ms post-stimulus-onset), and tapered off at 640 ms post stimulus onset. Peak competition decoding accuracy occurred at 220 ms post-stimulus-onset. For subsequent analysis, we chose the segment of time for which decoding was significant and for which accuracy was greater than one standard deviation (SD) above chance. This encompasses the time window from 80–600 ms post-stimulus-onset (note that accuracy dips below 1-SD-above-chance for a brief period between 100 to 200 ms; we decided to keep these timepoints in the analysis rather than introducing a discontinuity in the time window of interest).
To evaluate whether the neural representation of competition remained constant throughout the significantly decodable period, we computed a representational dissimilarity matrix (RDM)23 using the learned weight vectors at each time point. The result is shown in Fig. 5a. Over time, there seemed to be two distinct neural representations that were predictive: one early, from 100–200 ms post-stimulus-onset, and one late, from 300–600 ms post-stimulus-onset. We examined the difference between these representations by transforming the weight vectors into pattern maps using the technique described in Haufe et al.24. We then averaged these maps within each of the two time periods of interest and made topographic plots. The results are shown in Fig. 5b.
Predicting Subsequent Memory from the Swahili Learning Task
For each competition-localizer time point from 80 ms to 600 ms post stimulus onset, we re-trained the classifier on all of the competition data from Session 1 and applied it to the EEG data from Session 2, again within-subject and in a sliding window fashion, generating a grid of competition values for each retrieval practice round (where the grid is defined by which time points post-stimulus-onset were used for training vs. testing). For each item, we computed a measure of competition drop by subtracting the item’s competition value for the last round (R4) from the competition value for the round for which the participant first answered correctly. We then used this measure of competition drop to attempt to distinguish between items that were subsequently remembered vs. forgotten in Session 3 (this analysis was run separately for each grid point). The initial comparison is shown in Fig. 6. We had no a priori hypothesis about when in time our effect would occur, so we computed this grid search and then corrected for multiple comparisons over the grid.
Figure 6a,b show the result of predicting subsequent memory using the competition drop metric. Results that are consistent with our hypothesis that a larger competition drop predicts successful learning are shown in yellow. Results consistent with the opposite effect are shown in blue. Note the large swath of timepoints consistent with our hypothesis. In general, classifiers trained on early Session 1 time points (i.e., the bottom half of the grid) were most predictive of subsequent memory performance.
We determined the (multiple-comparisons-corrected) significance level of these results using a 2-stage process. We first computed an (uncorrected) p value for each grid square by comparing the competition drop scores for subsequently-remembered vs. subsequently-forgotten items using a t-test (for this test, we pooled all items across all subjects). Next, we thresholded the p values from this t-test at p < 0.10. To correct for multiple comparisons, we used a cluster permutation test. For this test, we computed the size of the largest cluster of contiguous grid squares where p < 0.10; we then compared this to a null distribution of maximum cluster sizes that we generated by permuting the assignment of items to the subsequently-remembered vs. forgotten classes; this analysis yields a corrected family-wise error p value. The large central cluster was the only cluster to achieve a family-wise p < 0.05, (p = 0.003), marked with red xs in Fig. 6b. The histogram of cluster sizes is shown in Fig. 6c.
To establish the reliability of this cluster over the participant population, we computed a participant-level bootstrap (i.e., resampling participants with replacement) and recorded the proportion of bootstrap draws for which a given grid square was included in a significant cluster. The points in the large cluster were present in nearly all bootstrap draws, as shown in Fig. 6d.
As a control, the same analysis was performed on the study trials from Session 2. The purpose of this analysis was to determine whether the effect was recall-specific (and therefore truly measuring competition) as opposed to tracking something more generic. The result is shown in Supplementary Fig. S1. Computing our competition drop score using the study trials failed to predict Session 3 performance, as no clusters of timepoints survived our cluster permutation test (the smallest cluster p value was p = 0.48), and very few points in the grid were consistently predictive across subjects.
These results show that the competition drop measure was strongly predictive of subsequent memory. Specifically, items for which there was a large drop in competition level from the first correct round to the last round were much more likely to be subsequently remembered. However, it is unclear from the result shown in Fig. 6 whether the competition drop metric computed from the Session 2 EEG data provides additional predictive power over the Session 2 behavioral accuracy data alone. To address this, we re-ran the analysis on subsets of items where Session 2 behavioral accuracy is held constant. Based on Fig. 3b, we selected items with 0011 and 0111 response configurations because these subsets both contained a large number of items and yielded a relatively even split between subsequently remembered and forgotten items. The result is shown in Supplementary Fig. S2. The results split by behavior (panels b and c) are qualitatively similar to both the original result (panel a) and to each other. Family-wise significance was more difficult to achieve for these sub-analyses, given the drastic reduction in items (574 for 0011 and 858 for 0111, as compared with 1550 items total); the resulting family-wise p values were p = 0.01 for 0011 items and p = 0.17 for 0111 items. Having said this, it is striking how similar the results were for the 0011 and 0111 analyses, given that they were computed on two entirely distinct subsets of the data — these results can thus be considered to be an internal replication of our main result, modulo the power issues noted above.
To rule out the possibility that these competition drop results simply reflect decreased reaction times (RTs) across retrieval trials, we ran another control analysis. In this analysis, we computed a “reaction time drop” score, using the same procedure that we used to compute competition drop, except here we measured the decrease in RTs instead of the decrease in our EEG measure of competition (note that the experiment was not originally designed to measure RTs, and participants were not allowed to respond prior to 2 s, to ensure that our EEG signal was not contaminated by movement artifacts). We found that reaction time drop was not predictive of subsequent Session 3 recall (p = 0.83). Supplementary Fig. S3 shows descriptive statistics for reaction time data. Average reaction times in Session 2 (for correctly-recalled items) numerically decreased across retrieval trials and were numerically lower for items that were subsequently recalled vs. subsequently forgotten in Session 3. However, across-item variance in reaction times was very large relative to these differences in means, making Session 2 reaction time useless as a predictor of subsequent memory for specific items.
Given that our main analysis pooled trials across subjects, both within-subject and across-subject variance could have contributed to the observed relationship between competition drop in Session 2 and subsequent recall in Session 3. That is, it could be the case that subjects who had higher competition drop scores on average had higher levels of Session 3 recall on average; or it could be the case that, within individual subjects, items that had higher competition drop scores were recalled better in Session 3. To address this latter point, we z-scored the competition drop value within subject. This eliminated the influence of between-subject variance. The recomputed grids for all trials, as well as the 0011 and 0111 items (considered separately) are shown in Supplementary Fig. S4. Qualitatively, these results strongly resemble the original results, but the predictive power is lessened by z-scoring within subject. This suggests that both across-subject variance and within-subject contributed to our results.
Finally, prior work has shown that theta band power in the first 500 ms post stimulus onset can predict retrieval-induced forgetting22. The results described above were computed using EEG voltages as the input to the classifier; however, the same process can be applied to any set of features derived from the EEG data. To test the ability of theta band power to predict the benefit of retrieval practice, we repeated the above analyses using theta band power values (computed via the Hilbert transform) as the features in our analysis pipeline. We found that high- and low-competition trials in Session 1 could be distinguished with theta band power, shown in Supplementary Fig. S5. While there were several timepoints that achieved significance, classifier accuracy was numerically lower than when we used voltage features (see Fig. 4). For subsequent analysis, to facilitate comparison between voltage and theta results, we used the same timepoints for our theta analyses that we used in our voltage analyses (note that this is a superset of the discriminative theta band timepoints). Using these timepoints and testing on the Session 2 data, we did not recover a cluster of timepoints that significantly predicted Session 3 behavior; the results qualitatively resemble the results for voltage features but were not as robust.
In this study, we trained an EEG classifier to detect neural correlates of competitive retrieval (using a separate competition localizer phase), and we used this classifier to track competition across repeated retrieval practice attempts for Swahili-English paired associates. As predicted, we found that the decrease in competition across retrieval practice trials (during learning) was predictive of subsequent recall success one week later. We did not have a strong a priori hypothesis about which time windows (post-stimulus onset) would yield a competition readout that predicts subsequent memory, so we swept across a grid of possible time points and found a large cluster of time points (significant after multiple-comparisons correction) that yielded the predicted relationship whereby competition decrease predicts better memory; we did not find any reliable effects in the opposite direction. Furthermore we found that this predictive relationship was only present in the recall trials from Session 2. A competition drop score computed over study trials from Session 2 failed to predict Session 3 performance.
Importantly, and unsurprisingly, it was also possible to use participants’ behavior during learning (i.e., the number of times an item was retrieved correctly during retrieval practice trials) to predict memory on the final test one week later. To assess whether our EEG measure of competition drop provided incremental predictive power, beyond what one would get from behavior alone, we ran follow-up analyses where we matched behavioral performance during learning and demonstrated a very similar pattern of results, although significance values were lower (reflecting the decreased number of trials included in these behavior-matched analyses). Furthermore, we assessed whether the drop in Session 2 recall reaction times (computed analogously to competition drop) could predict Session 3 performance, and found that it could not. One caveat is that, because we wanted 2 s of clean EEG data free from movement artifacts, we required that participants wait 2 s before responding during each Session 2 recall trial. This likely had an effect on the recorded reaction time for each trial, and it is possible that reaction time drop would have been a more useful predictive metric had the delay not been a part of the experiment.
Note also that, based on the results reported here, we can not completely rule out the possibility that competition drop is tracking difficulty as opposed to competition per se, given that high-competition retrieval is intrinsically more difficult that low-competition retrieval, all other things being equal. Disentangling competition from difficulty will require further experimentation where difficulty is manipulated orthogonally to retrieval competition (e.g., by perceptually degrading retrieval cues)25.
Notwithstanding this point about difficulty, the results are consistent with our hypothesis that benefits of testing on long-term retention are due (at least in part) to reductions in competition8. These reductions in competition could arise from multiple mechanisms, e.g., weakening or inhibition of competing memories13,26,27,28, differentiation of competing memories from the target memory19,29,30,31,32, or even integration of competing memories32,33,34. Reduced competition might also owe, in part, to an updated contextual representation stemming from retrieval-based context reinstatement; this, in turn, could have the effect of restricting the search set and allowing for relatively unfettered access to the target material35,36. The results of this study do not arbitrate between these possibilities.
Furthermore, as noted in the Introduction, researchers have set forth a wide range of explanations for the testing effect other than competition reduction1,2. For example, searching for the correct response on a test leads to the elaboration of semantic associations through which the target material can later be accessed7,8. The important point here is that these alternative explanations are not mutually exclusive with the competition-reduction account. Our study was designed specifically to evaluate the competition account and not the other theories; hence, additional work is needed to assess how these different accounts fit together.
Another limitation of the present study is that we did not include a condition where participants simply restudied items instead of doing retrieval practice. Based on prior work with this Swahili-English word-pair-learning paradigm3, we expect that this condition would have yielded lower levels of competition (as measured by the classifier) and worse recall compared to the retrieval practice condition. Given that our goal was explaining variance within the retrieval practice condition, we decided to omit the restudy condition in order to maximize the number of retrieval practice trials (thereby maximizing our power for detecting within-condition effects).
Our results resemble those of Kuhl et al.20, who found (in a between-subjects fashion) that a decrease in a neural measure of competition (in their case: anterior cingulate activity) predicted memory performance. However, there is an important difference between our results and those of Kuhl et al. – while we found a relationship between competition reduction during retrieval practice and subsequent memory for the practiced items, Kuhl et al. found that decreased anterior cingulate activity was specifically correlated with competitor forgetting (RIF) but not improved memory retention for the practiced items20. This Kuhl et al. result exemplifies the widely-reported regularity that RIF of competing items tends not to be correlated with improved memory for practiced items13,17.
What, then, explains why competition reduction was related to improved memory for practiced items in our study? One potential explanation for this discrepancy is that most RIF studies (including Kuhl et al.) used a much shorter retrieval delay than we did: The Kuhl et al. experiment included the retrieval practice and final test phases as part of the same fMRI session, whereas we included a 1-week delay between retrieval practice and the final test. Other studies have found that the beneficial effects of testing on target memory are most evident after long delays of the sort that we used here1,3. Several mechanisms might be responsible for these delay effects: For example, offline consolidation during the 1-week delay might serve to amplify representational changes (differentiation or integration effects) that are set in motion during retrieval practice37; alternatively, competition reduction might be more important for successful recall when the target is weak (after a long delay) then when it is strong. Regardless of the mechanism, if adding a 1-week delay amplifies the memory benefits of retrieval practice, this might also make it easier to see a relationship between these benefits and competition reduction during learning.
Another relevant question is how our results relate to those of Hanslmayr et al.22, which we used as the basis for our “competition localizer” analyses. One difference between the studies is that our main analysis path (chosen a priori) used voltages instead of the theta-power features used by the Hanslmayr et al. study. After completing the analyses with voltage features, we re-ran the main analyses with theta-power features. The results of these analyses, shown in Supplementary Fig. S5, conceptually replicate the prior finding22 that theta-power features discriminate between high-competition and low-competition items. Competition drop computed using theta-power features during Session 2 did not significantly predict subsequent memory; the pattern of results was qualitatively similar to what was obtained for voltages, but less robust.
In summary: We demonstrated for the first time that decreases in competition (measured neurally) across successive retrieval practice trials are predictive of subsequent retention. This extends prior work by Kuhl and colleagues showing that decreases in competition across trials predict competitor forgetting20, by showing that decreases in competition also predict long-term target retention. From a theoretical perspective, this work supports the hypothesis that competition decreases are (at least in part) responsible for the memory benefits of testing. Practically speaking, this result points to potential brain-computer-interface applications: By tracking the decrease in measured competition over time, we may be able to provide online feedback to learners about when a memory has been properly learned. More concretely, learners could be instructed to keep studying a pair until competition levels “bottom out,” indicating that competitors have been suitably neutralized. At present, the EEG signal is still too noisy to provide useful feedback to learners on a trial-by-trial basis, but future improvements in signal-to-noise may make this possible.
We collected data from 49 participants in total (32 female). Of these, 41 completed all three experimental sessions, while 8 completed only Session 1. Forty (25 female) of these individuals contributed sufficient Session 2 trials (after artifact removal) for subsequent analysis.
Participants were all right-handed native English speakers with normal or corrected-to-normal vision who reported no prior knowledge of Swahili. This study was approved by the Princeton University Institutional Review Board. All participants provided written, informed consent and were monetarily compensated for their time. All methods were performed in accordance with the relevant guidelines and regulations from the Princeton University Institutional Review Board.
Session 1: Competition Localizer Task
Session 1 utilized 19 categories, each of which was linked to eight exemplars of that category (both the category and exemplars were in English). These materials were largely based on the category-exemplar sets used by Anderson, Bjork and Bjork38. Exemplars within each category were selected to begin with a unique two-letter string, allowing for selective cueing. The full stimulus set can be found in Supplementary Table S1.
Session 2: Swahili Learning Task
For Session 2, we used the 40 Swahili-English word pairs from Karpicke and Roediger3, augmented with 20 similar Swahili-English word pairs. Pilot studies revealed that participants’ recall performance was nearly perfect when we used only 40 word pairs, leading to a lack of examples of forgotten items for subsequent analysis, hence the increase in number of items from 40 to 60. Stimuli can be found in Supplementary Table S2.
Session 3: Swahili Follow-up Test
In Session 3, participants were tested on the Swahili words presented in Session 2 using Google Forms.
A Biosemi ActiveTwo EEG Data collection system was used, in a shielded room featuring an RF/EMI Faraday cage built by Universal Shielding Company. The acquisition software was ActiView 7.07 on a 64-bit Windows workstation. A 64-channel configuration and a sampling rate of 512 Hz was used, along with 6 flathead electrodes (2 to record eye blinks, 2 for bilateral eye movement, and 2 on each mastoid for referencing).
EEG data were preprocessed using EEGLab39. First, the data were visually inspected for major artifacts and those sections of data removed. Then the data were band passed between 2 and 200 Hz, with a notch filter applied at 60 Hz to remove line noise. The electrodes were re-referenced using the whole scalp, then the data were parsed into epochs using stimuli onset and the epochs were baseline corrected. To remove eye-blink artifacts, the top two independent components were removed, after being visually inspected. After preprocessing, 2170 trials out of a total of 2400 (pooled across participants) from Session 2 survived for subsequent analysis. On average 295 out of 304 trials (per subject) from Session 1 survived.
For the experiment using theta band power, theta band was defined as between 4 and 8 Hz, inclusive, and was extracted using a Hilbert transform after the electrode referencing step. The data were then parsed into epochs without baseline correction or blink removal.
There were three experimental sessions: See Fig. 1 for a schematic. In Session 1, participants were shown the high/low competition retrieval stimuli while EEG was recorded. In Session 2, participants were given four study and test rounds for 60 novel Swahili words while EEG was recorded. Session 3 consisted of a cued-recall test of the Swahili words studied in Session 2. No EEG recording took place in Session 3.
Session 1: Competition Localizer Task
Session 1 was composed of two parts: an initial study phase and then a test phase. In the study phase, the participant was shown all the category-exemplar pairs in the experiment in one block. Each pair was presented for 2 s; a fixation cross was presented for 1 s between each pair. The category was presented above fixation and the exemplar was presented below fixation. Participants were instructed to try to remember the category-exemplar pairs. The test phase of the experiment consisted of four blocks in which the participant was presented with part of a category-exemplar pair. In each trial, the participant was presented with either the category or the exemplar. The other component of the pair was only partially presented, with the first two letters followed by two dashes (not diagnostic of the length of the word). Each partial pair was presented for 3 s, with 1 s of fixation between trials. Participants were instructed to silently fill in the blank with the correct word. Trials in which the category was provided required retrieval of the exemplar from many similar competitors, and are designated “high-competition trials”. Trials in which the exemplar was provided were assumed to reflect a relatively low level of retrieval competition and are referred to as “low-competition trials”. Participants were given the opportunity to rest between each block, as well as after the study block. Each block was composed of equal numbers of high-competition and low-competition trials, randomly interleaved. All 19 categories × 8 items × 2 conditions possible partial pairs were each shown once in the test phase.
Session 2: Swahili Learning Task
If time permitted, Session 2 was run immediately following Session 1; if not, participants returned within one week to complete Session 2. This session consisted of four rounds, with each round consisting of a study block and test block. In the study blocks, the participant read the Swahili-English word pairs in a random order. Each pair was presented for 2 s, and a fixation cross was presented for 1 s between trials. In the test blocks, participants were shown just the Swahili words (in a re-randomized order) and asked to respond with the English translation. Participants were asked to think of the answer silently for 2 s prior to typing it; a delay was programmed such that keystrokes occurring within 2 s of the Swahili word presentation would not appear. When finished, the participants pressed “enter” to move to the next Swahili word, and were free to skip words if need be. They were instructed not to spend too much time worrying about the response and to skip the question if unsure. Reaction time during the recall phase was operationalized as the time at which the first keystroke was detected, after the 2 s delay had elapsed. This keystroke could either be the start of a word or it could be the subject hitting ‘enter’ to skip the trial. Participants were given the opportunity to rest between each block. Subjects were asked not to review Swahili vocabulary between Session 2 and Session 3.
Session 3: Swahili Follow-up Test
One week after Session 2, participants returned to the same room to take a final test on the Swahili words presented in the previous session. Participants took the test in the same chair and with the same computer setup as the previous sessions, but they were not outfitted with an EEG cap as no electrophysiological data were recorded during this session.
Behavioral Data Analysis
For each Swahili-English word pair, each participant produced 4 behavioral datapoints in Session 2 (correct or incorrect recall, for each test round). We examined whether the response pattern of each individual item over the four rounds of Session 2 could predict which items were remembered or forgotten in Session 3. To evaluate the predictive power of the Session 2 responses, we used L2 penalized logistic regression. Each item was represented by a four dimensional binary vector indicating whether the participant answered correctly in each of the Session 2 rounds. These were the features used to predict the label, a binary scalar indicating whether the item was remembered in Session 3. To establish significance, a permutation test was conducted with 1000 permutations. In each permutation the item labels (subsequently remembered and subsequently forgotten in Session 3) were randomly shuffled before cross-validation, and then the classifier was trained and tested as usual.
EEG Predictive Modeling Framework
Our modeling framework can be outlined as follows: First, we located the times relative to stimulus onset in Session 1 at which high-competition and low-competition neural states were significantly distinguishable by a logistic regression classifier, with a suitably large effect size (accuracy >1 SD above chance). Then we trained our classifier on the data from each of those time points separately and applied it to each time point post-stimulus-onset, for each retrieval trial in Session 2. This created, for each of the four testing rounds for a given item, a grid of competition estimates (where the grid is defined by the time point where the classifier was trained in Session 1, and the time point where the classifier was applied in Session 2). We computed a competition drop score for each item by subtracting the competition estimate for the last round from the competition estimate for the first round in which the participant answered correctly; this was done separately for every grid square. We then used a t-test to establish, for each grid square, whether competition drop scores in that grid square were predictive of subsequent memory, and corrected for multiple comparisons. Each of these steps is described in more detail below. This framework was applied both to voltage data and to theta band power data.
Classifying High-Competition and Low-Competition Neural States
For each participant, we had Session 1 EEG data corresponding to roughly 300 trials (depending on visual artifact removal), with half of these trials consisting of high-competition retrieval and half consisting of low-competition retrieval. The EEG data consisted of a multidimensional timeseries of voltage (or theta-band power) at each channel and each timepoint. For each time point, we averaged the points for the next 50 ms, creating a 64-dimensional feature vector representing the average activity in each channel for the 50 ms period starting at that time point. This feature vector was input into a binary L2 penalized logistic regression classifier. The data were divided into five cross-validation folds. For each fold and time point, the classifier was trained on four of the folds and tested on the fifth. Thus we generated time series of classification accuracy for decoding high vs low competition states from the EEG data. This process was applied separately for each participant, and the accuracy timeseries was averaged across participants.
Significance and population-level reliability were measured using a within-participants permutation test with 100 permutations, and the within-participants p values were combined by treating participant as a random effect. Each participant’s p value was converted into a z value via the inverse normal transform. A group p value was generated by a one sample t-test between these z values and zero. The group p value at each timepoint was corrected for multiple comparisons using Bonferroni correction for an initial group p value of 0.001 divided by the total number of timepoints (55).
To generate the representational dissimilarity matrix over the weight vectors, we computed the cosine distance of the weight vectors for each pair of timepoints. To examine the neural representation underlying a given weight vector, we used the method described by Haufe et al.24, and averaged the resulting patterns over subjects and over the timepoints of interest.
Tracking Competition Level during Cued Recall
For each participant, a logistic regression classifier was trained to distinguish high-competition from low-competition trials using all trials from Session 1; a separate classifier was trained for each timepoint that was significant (and where accuracy was greater than 1 SD above chance) in the Session 1 group analysis. Each participant’s Session 2 data contained four retrieval rounds for each of the 60 words. These data, like the Session 1 data, form a multidimensional timeseries, and we again looked at each 50 ms window average separately, creating a 64-dimensional feature vector to which the trained classifier could be applied. Logistic regression output, when not thresholded into classes, can be treated as the estimated probability of a sample belonging to the “1” class. We refer to this output as the “competition probability,” the probability that high-competition retrieval is occurring. For each of the four retrieval rounds and each of the 60 items, our procedure generated a grid of this competition probability, with each point in the grid consisting of a training time from Session 1 and a test time from Session 2.
Predicting Successful Learning
The main goal of our analysis was to assess whether the decrease in competition probability across Session 2 rounds predicted subsequent recall in Session 3. To achieve this goal, we first computed (for each item, and for each grid square) a competition drop score: the competition probability for the first round in which the participant responded correctly minus the competition probability for the last round (R4). If the participant never responded correctly or did not respond correctly in round 4, the item was dropped from further analysis. This reduced the total number of trials from 2170 to 1550. In two separate analyses we further restricted the trials to those with behavioral patterns 0011 (574 trials) and 0111 (858 trials). For each grid square, we took the competition drop scores for all trials (pooled across all subjects) where the item was subsequently remembered, and compared them to the competition drop scores for all trials where the item was subsequently forgotten. The distributions of competition drop scores for subsequently-remembered and subsequently-forgotten items were Gaussian according to the Kolmogorov-Smirnov test40. Thus, we could use a one-tailed t-test to measure how well the competition drop distinguishes remembered and forgotten items. This generated a grid of p values for every grid square. Note that, unto themselves, the p values for this test tend to be liberal (insofar as it is a fixed-effects test), but this is remedied by the cluster-correction procedure described below.
To correct for multiple comparisons, we used a variant of the cluster permutation test to compute family-wise error41,42. First, we thresholded the p values at 0.10. We then computed the sizes of all clusters in the grid, where a cluster was defined as a set of contiguous grid squares that all had p values less than 0.10, and the cluster size was the number of grid squares belonging to the cluster. Contiguous was defined as immediately adjacent timepoints, not counting diagonal adjacency.
We then permuted the assignment of items to the subsequently-remembered and subsequently-forgotten clusters, re-computed the grid of p values, and thresholded again. For each permutation, the size of the maximum resulting cluster was stored. This was repeated 1000 times. The true clusters were then compared to the histogram of permuted cluster sizes and a family-wise corrected p value was generated. Only clusters with a family-wise p value less than 0.05 were considered significant.
To measure the population-level reliability of the surviving cluster, we sampled participants with replacement and recomputed the test statistic over all the timepoints. We then looked at the fraction of bootstrap draws in which each timepoint was part of a significant cluster. If the effect is present throughout the population, this should be close to 1 for all points in the true cluster. If the effect is instead dominated by only a few participants, the fraction would be much lower. We computed 1000 bootstrap draws.
We ran a control version of the above analysis using the study trials instead of the retrieval trials, to determine whether the measured effect was recall-specific. All criteria for trial selection were held constant, so the exact same trials for each item were used.
As another control analysis, we used reaction times from the Session 2 recall rounds to try to predict Session 3 performance. To that end, we computed the difference in reaction time from the first correct trial to the last trial for a given item, much in the same way that we computed the competition drop score. We then attempted to predict Session 3 performance using this reaction time drop score.
Data Availability Statement
The datasets generated during and analyzed during the current study are available via the Princeton University dataspace: http://arks.princeton.edu/ark:/88435/dsp015138jh55f. Code used for data collection and analysis are available at https://github.com/PrincetonCompMemLab/comp-eeg. Code used to generate Fig. 6 and Supplementary Figs S1, S2, S4 and S5 is available for manipulation via CodeOcean: https://doi.org/10.24433/CO.3032d861-7539-45c9-b83d-59f8c8ea6a15.v4
Publisher's note: Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.
This work was supported by the National Science Foundation Graduate Research Fellowship under Grant No. DGE1252522, and National Institutes of Health grants No. R01HD075328 and No. R01MH069456.