Decoding the tradeoff between encoding and retrieval to predict memory for overlapping events

Abstract

When new events overlap with past events, there is a natural tradeoff between encoding the new event and retrieving the past event. Given the ubiquity of overlap among memories, this tradeoff between memory encoding and retrieval is of central importance to computational models of episodic memory (O'Reilly & McClelland 1994; Hasselmo 2005). However, prior studies have not directly linked neural markers of encoding/retrieval tradeoffs to behavioral measures of how overlapping events are remembered. Here, by decoding patterns of scalp electroencephalography (EEG) from male and female human subjects, we show that tradeoffs between encoding and retrieval states are reflected in distributed patterns of neural activity and, critically, these neural tradeoffs predict how overlapping events will later be remembered. Namely, new events that overlapped with past events were more likely to be subsequently remembered if neural patterns were biased toward a memory encoding state-or, conversely, away from a retrieval state. Additionally, we show that neural markers of encoding vs. retrieval states are surprisingly independent from previously-described EEG predictors of subsequent memory. Instead, we demonstrate that previously-described EEG predictors of subsequent memory are better explained by task engagement than by memory encoding, per se. Collectively, our findings provide important insight into how the memory system balances memory encoding and retrieval states and, more generally, into the neural mechanisms that support successful memory formation.

Keywords: EEG; Encoding; Episodic memory; MVPA; Retrieval; Subsequent memory effect.

Figure 1.. Task Design and Behavioral Results.

(A) During List 1, subjects studied individual objects (e.g. bench, fan). During List 2, subjects saw novel objects that were from the same categories as the items shown in List 1 (e.g., a new bench, a new fan). Preceding each List 2 object was an “OLD” instruction cue or “NEW” instruction cue. The “OLD” cue signaled that subjects were to retrieve the corresponding item from List 1 (e.g., the old fan). The “NEW” cue signaled that subjects were to encode the current item (e.g., the new bench). Colored boxes are shown here for illustrative purposes and were not present during the actual experiment. Each run of the experiment contained a List 1 and List 2; object categories (e.g., bench) were not repeated across runs. After eight runs, subjects completed a two alternative force choice recognition test that tested memory for each List 1 and List 2 object. On each trial, a previously presented object, either from List 1 or List 2, was shown alongside a novel lure from the same category. The subject’s task was to choose the previously presented object. List 1 and List 2 objects were never presented together. (B) Behavioral results. Recognition accuracy is shown separated by list (1,2) and instruction condition (encode, orange; retrieve, teal). There was a significant interaction between list and instruction, primarily driven by greater accuracy for List 2 items presented with an encode instruction relative to a retrieve instruction. Error bars denote SEM; ** p < 0.01.

Figure 2.. Decoding memory states.

(A) We trained subject-specific L2-logistic regression classifiers to discriminate encode vs. retrieve trials during List 2. The classifiers were trained and tested on average spectral power across the 0-2000 ms stimulus interval with all electrodes and frequencies used as features. Mean classification accuracy across all subjects (solid vertical line) is shown along with a histogram of mean classification accuracies for individual subjects (black bars) and mean classification accuracy for permuted data across all subjects (dashed vertical line). Mean classification accuracy for permuted data ranged from 49.79% to to 50.31% across individual subjects (1000 permutations per subject). (B) Time-course of encoding evidence across the 2000 ms stimulus interval (i.e., the time window when the object image was on screen). Here, the classifier was trained on the full 2000 ms interval, as described in (A), but tested on 100 ms time windows. (C) Mean spectrogram of differences in spectral power for encode vs. retrieve trials as a function of electrode (y-axis) and frequency (x-axis). Orange indicates greater power for encode trials, teal indicates greater power for retrieve trials. Spectrograms were generated for each subject and then averaged across subjects. Bar graph below the spectrogram illustrates the mean spectral difference, averaging across electrodes and then across subjects, between encode vs. retrieve trials at each frequency. Error bars denote SEM. (D) Subject-specific logistic regression analyses tested whether trial-level encoding evidence derived from the classifiers during List 2 predicted accuracy on the subsequent recognition memory test. Separate regressions were performed to predict memory for List 1 items and List 2 items. Box and whisker plots show a positive relationship between encoding evidence during List 2 trials and subsequent memory for List 2 items but no relationship between encoding evidence during List 2 trials and subsequent memory for List 1 items. * p < 0.05, ** p < 0.01

Figure 3.. List 1 Univariate Subsequent Memory Effects.

(A) Mean spectrogram shows differences in spectral power for remembered vs. forgotten List 1 objects as a function of electrode (y-axis) and frequency (x-axis). Red indicates greater power for subsequently remembered items, blue indicates greater power for subsequently forgotten items. Spectrograms were generated for each subject and then averaged across subjects. Electrode names in bold text are the five electrodes that exhibited a reliable effect of frequency band (HFA vs. LFA; p < 0.01). These electrodes served as a functional region of interest (ROI) for subsequent analyses. (B) Subsequent remembering was associated with decreases in low frequency activity (LFA, < 28 Hz) and increases in high frequency activity (HFA, > 28 Hz), consistent with previous findings. Error bars denote SEM.

Figure 4.. List 2 encode/retrieve and List 1 SME comparison.

(A) The difference in spectral power between List 2 Encode and List 2 Retrieve trials in the functional ROI, separately for HFA and LFA bands. Error bars denote SEM. (B) Correlation between List 1 SME and List 2 encode/retrieve contrast. For each subject, we correlated the instruction contrast (encode - retrieve) and the subsequent memory contrast (remember - forget) at each electrode and frequency. The left and middle spectrograms illustrate this procedure. The right panel shows a histogram of zRho values across subjects. The average zRho value did not reliably differ from zero (t₃₅ = −0.7792, p = 0.4411).

Figure 5.. List 2 Univariate Subsequent Memory Effects.

Subsequent memory effects for the functional ROI from Figure 3. Each title describes the condition from which the EEG data were drawn (List 2 encode trials or List 2 retrieve trials) and the items from the recognition test that are included in the subsequent memory analysis (List 1 or List 2 items; note: the schematic shown in (A) also illustrates these relationships). For each plot in (B-D), each line reflects data from one of the five electrodes from the functional ROI. Subsequent memory effects for List 2 items significantly differed for encode vs. retrieve trials (panel C compared to panel D; p = 0.0300). Namely, when the goal was to encode (C), subsequent memory was predicted by relative HFA increases and LFA decreases, qualitatively identical to the pattern for List 1 items shown in (A). However, when the goal was to retrieve (D), a qualitatively opposite pattern was observed, with relative decreases in HFA and increases in LFA. Strikingly, for List 2 retrieve trials, HFA increases and LFA decreases predicted subsequent memory for to-be-retrieved List 1 items (E), similar to the canonical subsequent memory pattern as shown in (A). Thus, on retrieve trials, HFA increases and LFA decreases predicted relatively worse memory for the new List 2 item, but relatively better memory for the old List 1 item. Error bars denote SEM.