Separate memory-enhancing effects of reward and strategic encoding

Participants

A total of 24 young adults enrolled in the study, 4 of which were subsequently excluded from all analyses due to technical problems (n=1) or excessive head motion during scanning (n=3). Individuals excluded for excessive head motion had at least 3 runs with slice-averaged temporal signal to noise ratio (tSNR; mean signal across the time series divided by its standard deviation) < 60, while no other participant had more than 1 such run; poor image quality was confirmed by visual inspection of the data. The 20 included participants (13 F, 7 M) ranged in age from 18 to 39 (M_age = 26.8 years, SD = 5.9). All reported being right-handed, fluent English speakers, with no history of major neurological or psychiatric disorders, no current psychoactive medications, no color-blindness, and no other factors that would contraindicate MRI scanning.

All procedures were approved by the Northwestern University Institutional Review Board (IRB). Written consent was obtained from all participants. Participants were paid $20/hour; a typical session lasted 3 hours, with 2 hours in the MRI scanner. Participants were recruited via fliers on the Northwestern University Chicago campus, and via word of mouth.

Behavioral procedures and task stimuli

Participants were informed that they were participating in a memory study where the stimuli to be remembered were worth differing amounts of points that could be gained by accurate memory at test (points were not related to compensation or any extrinsic reward). Each study trial began with simultaneous presentation of two coin-shaped cues indicating the location and upcoming value of a stimulus on a 2x2 grid. High-value stimuli were worth 10 or 12 points, while low-value stimuli were rewarded with 2 or 3 points (Figure 1). Stimuli to be remembered were abstract visual “kaleidoscope” images generated using an algorithm initially described by Miyashita, Higuchi, Sakai, and Masui (1991) for creating novel, arbitrary, visual images by random deflection of colored polygons. On each trial, the value information was presented for 2.5s, and after a 1.0s blank delay, the two memory stimuli were presented on-screen simultaneously for 5 seconds, followed by another 0.5 s delay. After studying the visual stimuli, participants completed a baseline task consisting of a left/right (< or >) arrow direction judgment for 2s to 8s (times selected to optimize subsequent trial deconvolution), with each arrow appearing for 0.8 s each (plus 0.2 s ISI). An additional 1.0-sec ITI followed the final arrow. The arrows task was intended to keep participants’ attention focused on a low-level task during the baseline period to maximize effectiveness of contrasts among trial types (Stark & Squire, 2001). At the end of each scanning run of the encoding task, on-screen feedback was provided about accuracy on the arrows task for that list to encourage compliance with the baseline task.

Visual stimuli to be remembered were presented in blocks of 16 trials containing 2 stimuli (one high value, one low value) in each trial. Stimuli were presented twice each, always in the same quadrant and paired with the same value, but paired with different images on each presentation. Re-pairing was intended to prevent creation of item-item associations that could interfere with prioritization of high-value items. Each scanning run included 16 unique stimuli. Assignment of specific images to values was counterbalanced across participants. Across 6 runs, a total of 96 unique kaleidoscope images were studied in the scanner during the encoding phase.

After each study run, participants completed a memory test for the stimuli presented in the prior list. The memory test was used to post-hoc sort the successful and unsuccessful encoding trials. It was administered in the scanner and fMRI data were collected; these data will be reported in a separate publication. The test included the 16 images that had appeared on the preceding study list and 8 dissimilar foil images, presented in randomized order (see Figure 1). During the 4s presentation, participants were required to judge both whether the stimulus was old or new, and if old, which quadrant it had appeared in during study. Participants were instructed they needed to remember the location of old images to earn the stated point value from the study phase. Items that were correctly identified as old but in the wrong quadrant, and correctly-identified new items, were both awarded 1 point. Confidence judgments were provided after the memory test response and a 1s delay, but were not used in the analysis of the imaging data collected during study. For “old” responses, a 3-point confidence rating was provided: 1-confident in both the item being “old” and in its location, 2-confident that the item was “old” but guessing about location, or 3-guessing on both. For “new” responses, participants were asked whether that response was 1-confident or 2-guess. Confidence judgments were also followed by a jittered fixation interval, ranging in length from 2-8 seconds, optimized for trial deconvolution.

To familiarize participants with the study/test protocol structure prior to scanning, participants were given an initial practice phase consisting of 4 encoding trials, with 8 images presented once each followed by a short practice test phase in which 9 images (6 old, 3 new) were presented. Participants then completed one full-length encoding list and one full test list. Prior work has shown that extended practice leads to more consistent strategy use later, i.e., in the scanner (Castel, 2008; Cohen et al., 2016). Neuroimaging data were then collected from six full study lists, each followed by a test. After all 6 study-test cycles, an additional forced-choice recognition test for all study stimuli was administered (reported separately). After the fMRI session, participants were debriefed to gain some insight about the strategies that participants used during learning.

Scanning procedure

T2*-weighted echoplanar (EPI) images sensitive to blood oxygenation level dependent (BOLD) contrast were collected using a 3-T Siemens Prisma MRI scanner at the Northwestern Center for Translational Imaging (CTI). For the study task, each run lasted 4 minutes 28 seconds, and 130 whole-brain volumes were collected (after 4 discarded volumes at the beginning). Each functional volume contained 56 interleaved slices, TR = 2,000 ms, TE = 25 ms, flip angle = 80°, slice thickness = 2.0 mm, in-plane resolution = 2.0 x 2.0 mm, matrix = 104 x 98, FOV = 208 mm x 192 mm, and no gap between slices. We also collected a high-resolution EPI-navigated structural scan, with the following parameters: TR = 2,170 ms, TE = 1.69 ms, flip angle = 7°, 1 mm³ voxels, FOV = 176 mm x 256 mm x 256 mm, with GRAPPA acceleration. To minimize head movement during scanning, we placed cushions between the participant’s head and the coil. Stimuli were presented using PsychoPy v1.82 software (Peirce, 2007), and images were shown via a high-resolution MRI-compatible monitor (Nordic Neuro Lab, Milwaukee, WI), visible via a mirror placed on top of the head coil. Responses were collected using a 5-button fiber optic input device, connected to a response box in the scanner control room (MRA, Inc., Washington, PA) running a Cedrus RB-834 circuit board. The response pad interfaced with PsychoPy using the Cedrus PyXID driver library.

fMRI data analysis

Behavioral results

Participants exhibited better memory for high-value than low-value stimuli on the item and source/quadrant memory test (high-value items, M = 55.2% correct, SE = 4.0%; and low-value items, M = 43.4 % correct, SE = 4.4%), t(19) = 4.41, p < .001, d = 0.99. The rate at which old items were correctly judged as old regardless of the accuracy of the quadrant response (i.e., item hit rate) was also higher for high-value items (M = 88.9% correct, SE = 2.5%) than for low-value items (M = 81.6% correct, SE = 4.0%), t(19) = 2.95, p = .008, d = .66. Both high and low value items were judged old at a higher rate than new items on the test (i.e., the false alarm rate, M = 37.7%, SE = 4.0%), t(19) =13.08, p < .001, d = 2.93 (high-value items), and t(19) = 10.02, p < .001, d = 2.24 (low-value items). Memory within each value grouping (i.e., 2 vs. 3 points within low-value items and 10 vs. 12 points within high-value items) did not differ for either memory measure, all |t|s < 1.23, all p > .233.

Confidence judgments were well-aligned with response accuracy. Across images judged as old, the proportion for which participants reported being confident about both item and location was greater for trials in which both aspects were correctly remembered (M = 65.4%, SE = 4.1%) than for trials in which only the item judgment was accurate (M = 22.0%, SE = 3.6%), t(19) = 12.11, p < .001, d = 2.71, and was still lower for foils (M = 9.5%, SE = 2.7%), t(19) = 3.72, p = .001, d = .83. The proportion of images for which participants reported being confident only in the item recognition judgment was highest when only the item was correct (M = 56.5%, SE = 3.8%), was lower for foils (M = 41.6%, SE = 5.7%), t(19) = 3.55, p = .002, d = .79, and was still lower when both item and quadrant were correct (M = 24.6%, SE = 2.9%), t(19) = 2.63, p = .016, d = .59. Finally, judgments of an image as new were more likely to be made with confidence when the item was in fact a foil (M = 58.6%, SE = 6.0%) than when it was actually old (M = 42.6%, SE = 6.4%), t(17) = 3.52, p = .003, d = .83; two participants were excluded from this comparison because they had no missed old items with valid confidence responses. Because confidence was highly correlated with accuracy, it was not possible to separately incorporate confidence judgment accuracy into post hoc trial sorting.

On the post-study debriefing, participants described their primary encoding strategy as either associating images with a conceptual meaning or words (n=11), strategies related to perceptual features of the shapes within each image (n=8), or reported not using any particular strategy (n=1). Additionally, some participants described explicit efforts to focus more on high-value items (n=16), while others reported not using such efforts (n=4).

Brain regions related to successful encoding

A widespread set of brain networks exhibited greater activity for successful memory encoding than for stimuli that were not later successfully remembered (Figure 2; Table 1). Across all types of successful encoding trials, we found bilateral activity in frontoparietal regions such as dorsolateral prefrontal cortex and intraparietal sulcus, typically related to attentional control and working memory. There were also strong clusters of activation in ventral occipitotemporal regions typically associated with object and shape perception, such as lateral occipital complex (LOC), as well as some activation in more dorsal portions of lateral occipital cortex. Additional activity in more superior aspects of prefrontal cortex, in or near the frontal eye fields, was likely a function of the spatial nature of this memory task. Finally, when both items were successfully recalled, we found activity in brain regions typically associated with semantic encoding, such as left inferior prefrontal and left pre-supplementary motor area (pre-SMA). Brain activity was thus consistent with prior subsequent memory findings (cf., Kim, 2011).

Reward system activity and MTL-VTA connectivity.

Within the targeted reward system ROI (Figure 3A, see above for selection details), significantly increased activity was observed for successful encoding trials compared with unsuccessful memory (Figure 3B). This effect was observed when both items were recalled (H+L+), t(16) = 3.49, p = .003, d = .85, when only the high-value item was recalled (H+L−), t(16) = 2.53, p = .022, d = .61, and when only the low-value item was recalled (H−L+), t(16) = 2.65, p = .017, d = .64 (Figure 3B). A one-way repeated measures ANOVA showed no difference between these three conditions, F(2, 32) < 1, η_p² = .02. Even when individuals who reported not being explicitly selective are removed from the analysis, BOLD signal in the reward network did not differ between the 3 successful memory conditions, F(2, 24) < 1, η_p² = .06. Thus, evoked activity in the reward system appears to be associated with successful memory formation regardless of the value of information remembered.

To examine the selectivity of the successful memory effect to the reward ROI (and the areas shown in Figure 2), a parallel ROI analysis of the default mode network was carried out. The default mode network showed no evidence of differential activity during any type of successful encoding trial, relative to unsuccessful encoding, all |t|s < 1.44.

We then examined functional connectivity between a seed region in right MTL and VTA during successful versus unsuccessful encoding. A focus on R MTL is consistent with expectations from prior studies (see Methods). Initial analyses showed significantly greater connectivity between R MTL and VTA for the H+L+ and H−L+ conditions, and marginally greater connectivity for the H+L− condition (cluster .05 < p < .10), relative to H−L− trials (Table 2). These findings, and the observation (see Figure 3B) that univariate reward system activity was elevated to a similar degree with any successful encoding, motivated a combined second-level FEAT contrast combining together H+L+, H+L−, and H−L+ (any successful memory) trials, relative to H−L− trials (Figure 3C; see also Table 2). Parameter estimates within this cluster do not differ between the H+L+, H+L−, and H−L− conditions, F(2, 32) < 1, η_p² = .05. The degree to which memory success affected R MTL-VTA connectivity in the combined analysis correlated reliably with performance on the memory test, measured as mean proportion recall for all items (high-value and low-value), r = .58, p = .015 (Figure 3D). This relationship remained reliable after removal of the two outlier participants with substantially negative PPI values, r = .67, p = .007.

Selectivity analyses

None of the preceding analyses identified neural activity associated with the behavioral result showing better memory for high-value images. To strengthen the focus on top-down strategies, analyses described in this section exclude individuals (n=4) who reported not explicitly trying to remember high-value items better.² The trial type most clearly demonstrating selective encoding is H+L−. In contrast, H+L+ trials could represent a failure to be selective, but we found no evidence for this interpretation, as no brain areas were more active during H+L− trials than during H+L+ trials. It thus seems more likely that in terms of selectivity, H+L+ trials reflect either a successful effort to encode both items, or simply good memory. Our primary contrast for examining selectivity was between H+L− and H−L+ trials, however, as both trial types yield memory for one item, but only the H+L− condition reflects successful prioritization. Analysis of differential activity between H+L− trials and H−L+ trials identified a reliable difference bilaterally in dorsal posterior parietal cortex (dPPC), as well as a smaller cluster in anterior left inferior frontal gyrus (Figure 4A: Table 3). These regions were more active for successful high-value memory and likely reflect neural activity associated with effective strategic memory.

The posterior parietal effect was hypothesized to reflect the role of top-down attention during successful encoding, following Uncapher et al. (2011). If so, enhanced functional connectivity between dPPC and lateral occipitotemporal cortex would be expected during successful encoding, particularly for high-value items. To test this hypothesis, connectivity analysis was used to identify regions that might be working in concert with dPPC to control memory encoding. Search space was restricted to inferior portions of lateral temporal and lateral occipital cortex, as described in Methods. A significant positive effect was found in right lateral temporal cortex when contrasting combined effects from H+L− and H+L+ trials against that from H−L− trials (Figure 4B; Table 4). The magnitude of the PPI effect shown in Figure 4B was found to be correlated with memory selectivity (the difference in the proportion of items recalled for high-value vs. low-value items), r = .68, p = .010 (Figure 4C). The total recall rate across all items was not reliably correlated with the PPI effect, r = .37, p = .21. An additional PPI activation was observed in left lateral temporal cortex when combining across H+L+, H+L−, and H−L− trials, relative to H−L− trials (Table 4). The magnitude of this effect showed a marginal correlation with memory selectivity, r = .53, p = .065, but no correlation with the total recall rate, r = .03, p = .92. Thus, although connectivity effects are not limited to trials in which a high-value item was successfully encoded, the overall strength of parietal-temporal connectivity during successful encoding appears to be relatively more associated with selectivity than with overall memory.