If the hippocampus and vmPFC are more involved in attentional behaviors that are guided by memory, then they should show enhanced univariate activity during the memory-guided vs. explicitly instructed task. To examine this, we compared BOLD activity in these regions during the attention task (i.e., when the images were on the screen and participants were attending to artistic style or room layout). Indeed, BOLD activity was higher for the memory-guided vs. explicitly instructed task in both hippocampus, t(28) = 2.54, p=0.017, d = 0.47, 95% CI [0.872, 8.125], and vmPFC, t(28) = 3.74, p=0.0008, d = 0.69, 95% CI [2.518, 8.611] (Figure 3A).

Figure 3

Download asset Open asset

To determine if this difference in univariate activity is related to differences in behavioral performance across tasks, we examined whether A’ differences on the memory-guided vs. explicitly instructed task predicted univariate activity differences between these two tasks, across individuals. This relationship was not statistically significant in hippocampus (R² = 0.03, p=0.39, 95% CI [−0.502, 0.214]) or in vmPFC (R² = 0.02, p=0.52, 95% CI [−0.470, 0.253]). Thus, univariate activity enhancement in these regions for memory-guided attention cannot be explained solely by differences in behavioral performance.

If the hippocampus and vmPFC work together to establish memory-guided attentional states, then the extent to which one region’s activity is modulated by memory-guided attention might predict how much the other region’s activity shows such modulation. Indeed, the activity enhancement in each region for memory-guided attention (i.e., the BOLD activity difference for memory-guided vs. explicitly instructed tasks) was strongly correlated across individuals, R² = 0.51, p=0.000022, 95% CI [0.477, 0.867] (Figure 3B). Importantly, this correlation remained significant when controlling for individual differences in behavioral performance across tasks (R² = 0.50, p<0.0001). Together, these results suggest that the hippocampus and vmPFC play a similar functional role in memory-guided attention, and may be working together. In the Discussion, we further consider what enhanced univariate activity in these regions might reflect.

Our primary question was whether the hippocampus and vmPFC can use memory to prepare for upcoming attentional states. Thus, we differentiate between two main periods on any given trial: (1) the image period, when images are on the screen and participants are actively attending to artistic style or room layout, and (2) the orienting period, when participants are pushing a button to initiate the trial and seeing the attentional cue (Figure 1). In the explicitly instructed task, participants simply choose which button to press, and then the attentional cue (‘ART’ or ‘ROOM’) is randomly assigned. In the memory-guided task, participants select ‘art’ or ‘room’ as the attentional state based on memory for the preceding trial. Based on prior studies showing that hippocampal univariate activity is enhanced in preparation for upcoming, memory-guided attentional goals (Stokes et al., 2012), we predicted that the hippocampus — and vmPFC, given their tight connection for memory-guided behavior (Euston et al., 2012; Kaplan et al., 2017; Shin and Jadhav, 2016) — would show preparatory coding during the orienting period. Specifically, we predicted that during the orienting period, activity patterns in the hippocampus and vmPFC would resemble the attentional state (i.e., art vs. room) that is upcoming in the image period, primarily when that attentional state was selected on the basis of memory.

In order to test this prediction, we first needed to establish that the hippocampus and vmPFC differentiate between the two attentional states (art vs. room) during the image period. This would then allow us to determine whether neural signatures of the art vs. room states appear in a preparatory fashion during the orienting period, particularly for memory-guided attention. Our past fMRI studies indicate that the hippocampus does indeed differentiate between the art vs. room states during the image period (Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b), but here we sought to replicate this and extend it to vmPFC.

To this end, we obtained patterns of activity in the hippocampus and vmPFC for each image period, and then correlated these activity patterns as a function of the participants’ attentional state (Figure 4A). We compared pattern similarity for trials of the same attentional state (i.e., art-art, room-room) to pattern similarity for trials of different attentional states (i.e., art-room) separately for the memory-guided and explicitly instructed tasks. If a brain region represents online attentional states, then pattern similarity should be higher for trials of the same state vs. trials of different states (Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b). This was the case for both the hippocampus and vmPFC, in both the memory-guided (hippocampus: t(28) = 3.82, p=0.00067, d = 0.71, 95% CI [0.003, 0.009], vmPFC: t(28) = 6.58, p<0.0001, d = 1.22, 95% CI [0.009, 0.017]) and explicitly instructed tasks (hippocampus: t(28) = 7.12, p<0.0001, d = 1.32, 95% CI [0.006, 0.011], vmPFC: t(28) = 6.07, p<0.0001, d = 1.13, 95% CI [0.010, 0.021]). These results confirm that the hippocampus and vmPFC represent online attentional states (Figure 4B), a necessary precursor for examining preparatory attentional states during the orienting period.

Figure 4

Download asset Open asset

Having confirmed that the hippocampus and vmPFC represent online attentional states (i.e., attentional states during the image period), we next tested whether these regions also represent preparatory attentional states — i.e., whether their activity patterns during the orienting period code for attentional states that are upcoming during the image period.

To that end, we first calculated ‘template’ patterns of activity by averaging activity patterns during the image period across trials, separately for the art and room attentional states (Figure 5A). These ‘template’ activity patterns indicate, for a given brain region, what the BOLD activity pattern looks like when participants are actively attending to artistic style vs. room layout in the 3D-rendered images. We then correlated activity patterns during each individual orienting period with the two templates, and binned these correlations based on whether the template matched the orienting period attentional cue (e.g., correlation between the art template and the orienting period activity pattern on an art trial) or mismatched (e.g., correlation between the room template and the orienting period activity pattern on an art trial). This was repeated for each trial, and the resulting correlations were averaged separately for the memory-guided and explicitly instructed tasks. Lastly, in order to obtain a measure of preparatory attentional state representations, we calculated the difference between match-to-same-template pattern similarity and match-to-different-template pattern similarity. If a brain region shows preparatory coding, its orienting period activity patterns should resemble the same-state template more than the different-state template.

Figure 5 with 2 supplements see all

Download asset Open asset

Indeed, for the memory-guided task in the hippocampus (Figure 5B), activity patterns during the orienting period resembled the upcoming attentional state more than the other attentional state, t(28) = 4.78, p=0.00005, d = 0.89, 95% CI [0.008, 0.021]. Unexpectedly, this effect was also observed when attention was explicitly instructed, t(28) = 2.71 p=0.011, d = 0.50, 95% CI [0.001, 0.007]. Critically, however, preparatory attentional states in the hippocampus were stronger for the memory-guided vs. explicitly instructed task, t(28) = 3.18, p=0.004, d = 0.59, 95% CI [0.004, 0.017].

In vmPFC, activity patterns during the orienting period resembled the upcoming attentional state more than the other attentional state for both the memory-guided, t(28) = 6.25, p<0.00001, d = 1.16, 95% CI [0.010, 0.019], and explicitly instructed tasks, t(28) = 4.12, p=0.00030, d = 0.77, 95% CI [0.007, 0.020]. Contrary to our hypothesis, this effect was not significantly different between the tasks, t(28) = 0.77, p=0.45, d = 0.14, 95% CI [−0.003, 0.006]. Thus, the hippocampus, but not vmPFC, preferentially represented upcoming memory-guided vs. explicitly instructed attentional states.

For the preceding analysis, we used common image period templates for the memory-guided and explicitly instructed tasks: Art trials from both tasks were used to create an art template, and room trials from both tasks were used to create a room template (Figure 1—figure supplement 1). This was done because using separate templates for each task might artificially create differences in orienting period pattern similarity values even if the orienting period patterns do not differ across tasks (e.g., different numbers of correct vs incorrect trials across tasks may lead to different template activity patterns). Furthermore, as in the image period analysis and our previous work (Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b), we only used valid trials for the image period templates (this was to prevent neural activity related to invalid probes from contaminating image period activity patterns). Finally, we analyzed all orienting periods in each task, whether the previous trial contained a stay/switch cue or did not contain one of these cues (i.e., ‘no-cue’ trials). (Note that in the explicitly instructed task, stay/switch cues were embedded in the search set but had no relevance for the attentional state on the following trial). We included no-cue trials because, in the memory-guided task, the attentional state decisions following these trials still had to be guided by memory: In order to know that the attentional goal could be chosen freely, participants needed to remember that no stay or switch cue was presented on the previous trial.

However, one could argue for alternatives to each of these decisions (Figure 1—figure supplement 1). For example, because the memory-guided and explicitly instructed tasks had different demands during the image period, it can be argued that separate image period templates should be used for each task. Furthermore, contamination of image period brain activity by invalid probes should not differ across the memory-guided and explicitly instructed tasks, so one could argue for including invalid trials as well. Finally, attentional-state decisions following trials in which no stay or switch cue was presented might be less memory-driven than those following a stay or switch cue. This is because memory for the particular type of cue was required for choosing the correct attentional state following stay and switch cue trials, but memory for the mere presence or absence of a cue was sufficient following no-cue trials. Thus, one could argue that orienting periods following ‘no cue’ trials should be excluded from analyses. We therefore tested the robustness of our orienting period results by re-running the analyses with these alternative decisions.

We replicated the same pattern of results when: (i) using separate image period templates for the memory-guided vs. explicitly instructed tasks as opposed to a common template; (ii) using image period templates that include both valid and invalid trials as opposed to valid trials only, and (iii) analyzing only those orienting periods that followed either a stay cue or a switch cue (i.e., excluding orienting periods following no-cue trials).

Specifically, for the former analysis (i), we replicated the finding of preparatory attentional states in both hippocampus (memory-guided: t(28) = 4.50, p=0.00011, d = 0.84, 95% CI [0.007, 0.019]; explicitly-instructed: t(28) = 2.26, p=0.032, d = 0.42, 95% CI [0.0003, 0.006]), and vmPFC (memory-guided: t(28) = 5.02, p=0.00003, d = 0.93, 95% CI [0.008, 0.018]; explicitly-instructed: t(28) = 3.79, p=0.00073, d = 0.70, 95% CI [0.006, 0.019]). As in the main analysis, these preparatory attentional states were stronger for the memory-guided vs. explicitly instructed task in the hippocampus, t(28) = 3.32, p=0.0025, d = 0.62, 95% CI [0.004, 0.016], but did not significantly differ across tasks in vmPFC, t(28) = 0.31, p=0.76, d = 0.06, 95% CI [−0.005, 0.006].

For the second analysis (ii), we also replicated the finding of preparatory attentional states in both hippocampus, (memory-guided: t(28) = 4.24, p=0.00022, d = 0.79, 95% CI [0.007, 0.020]; explicitly-instructed: t(28) = 2.69, p=0.012, d = 0.50, 95% CI [0.001, 0.009]), and vmPFC (memory-guided: t(28) = 6.29, p<0.00001, d = 1.17, 95% CI [0.009, 0.018]; explicitly-instructed: t(28) = 4.23, p=0.00023, d = 0.78, 95% CI [0.006, 0.018]). Once again, preparatory attentional states were stronger for the memory-guided task in the hippocampus, t(28) = 2.38, p=0.024, d = 0.44, 95% CI [0.001, 0.015], and did not significantly differ between tasks in vmPFC, t(28) = 0.74, p=0.47, d = 0.14, 95% CI [−0.003, 0.006].

Finally, for the third analysis (iii), we again replicated the finding of preparatory attentional states for both the memory-guided, t(28) = 6.56, p<0.00001, d = 1.22, 95% CI [0.011, 0.021], and explicitly instructed tasks in vmPFC, t(28) = 4.00, p=0.00042, d = 0.74, 95% CI [0.007, 0.021]. In hippocampus, preparatory attentional states were again present for the memory-guided task, t(28) = 4.14, p=0.00029, d = 0.77, 95% CI [0.008, 0.025], but failed to reach significance in the explicitly instructed task (t(28) = 1.84, p=0.076, d = 0.34, 95% CI [−0.0004, 0.007]; note that this analysis is reduced in power because 1/3 of the trials were dropped). Once again, preparatory attentional states were stronger for the memory-guided task in the hippocampus, t(28) = 3.16, p=0.0038, d = 0.59, 95% CI [0.005, 0.022], but did not significantly differ across tasks in vmPFC, t(28) = 0.70, p=0.49, d = 0.13, 95% CI [−0.004, 0.007].

Thus, the main results are robust to many different analysis decisions. However, there is another potential concern. Are the observed results due to autocorrelation between orienting period and image period activity patterns, as a result of sluggish hemodynamic signals? We believe not, for several reasons. First, autocorrelation between the orienting period and image period should be higher for the explicitly instructed vs. memory-guided task because response times to initiate the trial were on average shorter for the explicitly instructed task (0.92 s vs. 1.12 s; t(28) = 2.44, p=0.021, d = 0.45, 95% CI [0.032, 0.361]). However, preparatory coding was stronger for the memory-guided task in the hippocampus and did not differ between tasks in vmPFC. Second, the image period templates — against which orienting period activity patterns were compared — were obtained from different runs of the task to remove within-run autocorrelation (Mumford et al., 2014). Third, the last brain volume (TR) for the orienting period and the first brain volume for the image period were excluded from the analysis to reduce autocorrelation between the image period and orienting period signals. (Note that the last brain volume for the orienting period was not dropped if that was the only volume during which the attentional cue was presented). Thus, we argue that the preparatory attentional state representations observed in the orienting period are not simply the result of autocorrelation between orienting period and image period activity patterns.

One could argue that dropping the last brain volume for the orienting period activity pattern disadvantages the opportunity to detect preparatory coding for the explicitly instructed task more than the memory-guided task. This is because, for the memory-guided task, the attentional state for trial N+1 is known as soon as trial N is over; but for the explicitly instructed task, it is only known when the attentional cue is presented at the end of the orienting period. To confirm that this is not the case, we re-ran the orienting period analysis including the last orienting period brain volume, and obtained the same pattern of results. We observed preparatory attentional states for both the memory-guided task (hippocampus: t(28) = 4.84, p=0.00004, d = 0.90, 95% CI [0.008, 0.021], vmPFC: t(28) = 6.42, p<0.00001 d=1.19, 95% CI [0.010, 0.020]) and the explicitly instructed task (hippocampus: t(28) = 3.11, p=0.0043, d = 0.58, 95% CI [0.002, 0.008], vmPFC: t(28) = 4.18, p=0.00026, d = 0.78, 95% CI [0.007, 0.020]). Importantly, preparatory attentional states were stronger for the memory-guided vs. explicitly instructed tasks in the hippocampus, t(28) = 3.04, p=0.00504, d = 0.57, 95% CI [0.003, 0.017], and no difference in preparatory attentional states across tasks was measured in vmPFC, t(28) = 0.86, p=0.40, d = 0.16, 95% CI [−0.003, 0.006]. Together, these findings suggest that our results are robust and cannot be attributed to idiosyncratic analysis decisions.

We argue that multivariate patterns of activity in the hippocampus during the orienting period reflect preparation for upcoming attentional states. However, is it possible that these activity patterns instead reflect retrieval of the attentional state from the previous trial? This is unlikely for the explicitly instructed task, where memory for the previous trial is not relevant for the attentional state on the current trial. Thus, preparatory signals for the explicitly instructed task in the hippocampus and vmPFC likely index anticipation of the upcoming task rather than memory retrieval. For the memory-guided task, however, it is possible that participants use the orienting period of a given trial to retrieve what they did on the previous trial. For example, during the orienting period for an upcoming ‘room’ trial, a participant may remember that the previous trial was a ‘room’ trial with a stay cue (or an ‘art’ trial with a switch cue). Are hippocampal activity patterns reflecting such memory retrieval?

Trials in which participants stay in the same attentional state as the previous trial are ambiguous: Remembering the previous trial and preparing for the current trial would be indistinguishable with our analysis because the attentional states are the same. However, trials in which participants switch from one attentional state to the other provide a strong test of our hypothesis. If hippocampal activity patterns during the orienting period reflect memory retrieval of the previous trial, they should resemble the previous attentional state more than the upcoming attentional state. If, however, hippocampal activity patterns during the orienting period reflect preparation, they should resemble the upcoming attentional state more than the previous one.

Indeed, when we analyzed only the trials that followed a switch cue (Figure 5—figure supplement 1), we found that hippocampal activity patterns during the orienting periods of the memory-guided task resembled the upcoming attentional state more than the other (previous trial’s) attentional state, t(28) = 2.90, p=0.0072, d = 0.54, 95% CI [0.004, 0.023]. We also conducted this analysis for the explicitly instructed task for completeness (although, here, a switch cue has no relevance for the attentional state on the following trial). Here, we found no evidence for an attentional state representation (neither the upcoming attentional state nor the previous attentional state) during the orienting period, t(28) = 0.16, p=0.87, d = 0.03, 95% CI [−0.006, 0.007]. (We are cautious in over-interpreting this null effect because this analysis contains roughly one-third the trials in the main analysis, and hence has lower statistical power.) Finally, as in our main analysis, multivariate evidence for upcoming attentional states in the hippocampus was higher for the memory-guided vs. explicitly instructed task, t(28) = 2.85, p=0.008, d = 0.53, 95% CI [0.004, 0.023]. These results therefore suggest that, during the memory-guided task, hippocampal activity patterns during the orienting period reflect preparation for the upcoming attentional state rather than retrieval of the preceding attentional state. This preparation for upcoming attentional states may involve memory retrieval of task-relevant goals and/or the use of these memories to bias neural processing toward task-relevant features. We discuss the content of such preparatory signals in more detail in the Discussion.

For completeness, we also analyzed only those trials following a switch cue for vmPFC and replicated our main results: Activity patterns during the orienting period resembled the upcoming attentional state more than the other (previous trial’s) attentional state for both the memory-guided, t(28) = 4.29, p=0.00019, d = 0.80, 95% CI [0.009, 0.027] and explicitly instructed tasks, t(28) = 4.45, p<0.0001, d = 0.83, 95% CI [0.009, 0.023]. These preparatory states did not significantly differ across tasks, t(28) = 0.71, p=0.48, d = 0.13, 95% CI [−0.004, 0.008]. These results suggest that for vmPFC — as for hippocampus — activity patterns during the orienting period reflect preparation for the upcoming attentional state, rather than retrieval of the previous attentional state.

Our results so far indicate that the hippocampus is more strongly engaged by memory-guided vs. explicitly instructed attention (Figure 3) and represents both current (Figure 4) and upcoming (Figure 5) attentional states. Moreover, the hippocampus shows stronger preparation for memory-guided attention. How does the hippocampus transform memory cues in the environment (i.e., stay/switch cues) into preparatory attentional signals? One possibility is that hippocampal interactions with visual cortex are enhanced when memory must be used to guide attention. This would allow mnemonically relevant information in the environment to be detected via hippocampal-visual cortex communication. Once this information is detected, the hippocampus can then use it to prepare for attentional states that are guided by those mnemonic cues. To test this, we examined whether functional coupling between the hippocampus and visual cortex is enhanced for memory-guided attention. Because detection of stay/switch cues requires being in a task-relevant attentional state, we hypothesized that the attentional states of the hippocampus and visual cortex will be more strongly aligned for the memory-guided task.

To examine this, we capitalized on novel neuroimaging methods that allow investigation of multivariate coupling between regions: multivariate (or informational) connectivity (Aly and Turk-Browne, 2016b; Anzellotti and Coutanche, 2018; Coutanche and Thompson-Schill, 2013). We focused on visual areas V1-2 because representations in these regions are correlated with those in the hippocampus during memory retrieval and predictive coding (Bosch et al., 2014; Hindy et al., 2016).

We first determined the quality of attentional states in the hippocampus and V1-2 on a trial-by-trial basis. This was achieved by determining whether activity patterns on a given trial more strongly aligned with the task-relevant vs. task-irrelevant attentional state (e.g., on a trial with a ‘good’ room attentional state, hippocampal activity patterns should more strongly resemble the average room-state activity pattern vs. the average art-state activity pattern). We then correlated these measures of attentional state ‘quality’ across the hippocampus and V1-2.

Prior to measuring multivariate connectivity, we first had to confirm that V1-2 represents current attentional goals (a precursor to examining the covariation of attentional states between regions is that each region must represent attentional states; see Figure 4). Indeed, in V1-2, pattern similarity was higher for trials of the same attentional state vs. trials of different attentional states, for both memory-guided, t(28) = 9.32, p<0.0001, d = 1.73, 95% CI [0.092, 0.144], and explicitly instructed tasks, t(28) = 11.83, p<0.0001, d = 2.20, 95% CI [0.103, 0.146]. Next, we computed multivariate connectivity between the hippocampus and V1-2, as described above (Figure 6A). High multivariate connectivity (i.e., inter-regional correlation) indicates that when one region is in a ‘good’ attentional state, the other region is also in a good attentional state, and when one region is in a ‘bad’ attentional state, the other region is also in a bad attentional state. Multivariate connectivity was significantly above zero in the memory-guided task, t(28) = 4.28, p=0.00020, d = 0.80, 95% CI [0.061, 0.173], but not in the explicitly instructed task, t(28) = 1.78, p=0.086, d = 0.33, 95% CI [−0.009, 0.123]. The difference between tasks was statistically significant, t(28) = 2.28, p=0.030, d = 0.42, 95% CI [0.006, 0.114]. These findings raise the possibility that covariation in attentional states between the hippocampus and early visual cortex may enable mnemonically relevant information in the environment to be detected, and then acted upon, to guide behavior on the basis of memory (Figure 6B).

Figure 6

Download asset Open asset

We have focused on multivariate measures of preparatory coding in the hippocampus: The extent to which orienting period activity patterns contain information about upcoming attentional states. Yet, a previous study found univariate activity enhancements in the hippocampus when memory was used to prepare for upcoming attentional goals (Stokes et al., 2012). In that study, hippocampal activity was enhanced when information in memory was available about an upcoming target location, even prior to the onset of attentional search. Here, we found that hippocampal activity levels are enhanced for memory-guided vs. explicitly instructed attention during the image period (Figure 3), but to more closely parallel the Stokes et al. (2012) study, we also examined whether univariate activity is enhanced during the orienting period, i.e., in anticipation of the attentional search task. However, during the orienting period, univariate activity in the hippocampus was not significantly different for memory-guided (M = 14.144, 95% CI [9.123, 19.164]) vs. explicitly instructed attention (M = 17.740, 95% CI [11.077, 24.402]), t(28) = 1.28, p=0.21, d = 0.24, 95% CI [−9.369, 2.176]. For completeness, we also examined univariate activity in vmPFC during the orienting period, but again found no significant difference between the memory-guided (M = 12.044, 95% CI [6.219, 17.870]) and explicitly instructed tasks, (M = 14.563, 95% CI [8.447, 20.678]), t(28) = 0.85, p=0.40, d = 0.16, 95% CI [−8.593, 3.556]. We return to this difference between the results of our study and those of Stokes et al. (2012) in the Discussion.

Although our focus has been on the hippocampus and vmPFC, we conducted exploratory whole-brain analyses to investigate neural signatures of attentional preparation elsewhere in the brain. We used a searchlight approach to find brain regions whose orienting period activity patterns were significantly correlated with their image period activity patterns. This approach was used to look for regions that showed greater preparation for memory-guided vs explicitly instructed attention, and regions that showed preparatory coding for either task treated separately. No voxels survived correction for multiple comparisons (p<0.05 family-wise error corrected) when looking for regions that showed greater preparation for memory-guided vs explicitly instructed attention. When we looked for preparatory coding for each task separately, a few isolated voxels survived correction for multiple comparisons but no meaningful clusters emerged (Figure 5—figure supplement 2). These results must of course be treated with caution: it is very likely that brain areas other than the hippocampus and vmPFC prepare for upcoming attentional goals, but more targeted region-of-interest analyses are required to uncover them.

There were two attentional states (art, room) and two tasks (memory-guided, explicitly instructed; Figure 1). In the ‘art’ attentional state, participants had to attend to the style of the painting in the base image (use of color, brushstrokes, level of detail) and determine whether any of the paintings in the search set could have been painted by the same artist who painted the painting in the base image (i.e., an art match: a painting that is similar in style). In the ‘room’ attentional state, participants had to attend to the layout of the room in the base image (arrangement of furniture, angles of the walls), and determine whether any of the rooms in the search set had the same spatial layout from a different perspective (i.e., a room match). Other aspects of the rooms (e.g., wall color, specific furniture exemplars) differed between the base image and its room match.

In the explicitly instructed task, the attentional state (art or room) was randomly assigned on each trial. In the memory-guided task, participants used memory for learned stay/switch cues to select their attentional goals: A stay cue on trial N indicated that the participant should stay in the same attentional state on trial N+1, while a switch cue on trial N indicated that the participant should switch to the other attentional state on trial N+1 (e.g., switch from ‘room’ to ‘art’ or from ‘art’ to ‘room’). Finally, some trials contained neither a stay nor a switch cue. Following those ‘no-cue’ trials, participants were free to choose either ‘art’ or ‘room’ as their attentional state on the next trial.

Participants completed 4 runs of the memory-guided task and 4 runs of the explicitly instructed task (25 trials per run). All runs of the same type were completed before switching to the other task, and task order was counterbalanced across participants.

The images used in this study were 3D-rendered rooms, each of which contained one painting. The rooms were designed with Sweet Home 3D (sweethome3d.com). Each room contained multiple pieces of furniture and had a unique shape and layout. A second version of each room (to be used as its ‘room match’) was created with a 30° viewpoint rotation (half clockwise, half counterclockwise) and altered such that the content was different, but the spatial layout was the same. This was accomplished by changing the colors of the walls and replacing the furniture with different furniture of the same type at the same position (e.g., replacing a chair with another chair). The paintings were chosen from the Google Art Project. To obtain the ‘art match’ for each painting, a painting from the same artist was chosen, which had a similar style but whose content could differ. The combined images (art in a room) were generated by manually ‘hanging’ each painting along a wall.

2 paintings and 2 rooms were chosen to be ‘stay’ and ‘switch’ cues (1 painting and 1 room were ‘stay’ cues; 1 painting and 1 room were ‘switch’ cues). 12 ‘cue’ images were generated by pairing each art cue (1 stay cue and 1 switch cue) with 3 different rooms, and each room cue (1 stay cue and 1 switch cue) with 3 different paintings. Thus, each stay/switch cue could appear in 3 different images. The 3 room ‘backgrounds’ for the art stay cue were the same as the 3 room ‘backgrounds’ for the art switch cue. Likewise, the 3 paintings embedded in the room stay cue were the same as the 3 paintings embedded in the room switch cue. The task-irrelevant portion of each stay/switch cue (the room in art stay/switch cues and the art in room stay/switch cues) was therefore not diagnostic of the cue’s identity.

The stimulus set used in the fMRI scan session contained 141 unique images (129 main images plus the 12 stay/switch cue images). These were derived from a set of 120 images (Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b) that were created by pairing each of 40 rooms with 3 different paintings (all by different artists) and each of 40 paintings with 3 different rooms (all with a different layout). We modified this set in order to pair the task-irrelevant feature of each stay/switch cue (e.g., the art in a room stay cue, or the room in an art stay cue) with 2 images used in the main stimulus set. That is, each of the 3 room ‘backgrounds’ for art stay/switch cues was also paired with 2 different paintings from the main stimulus set, and each of the 3 paintings embedded in room stay/switch cues was also paired with 2 different rooms from the main stimulus set. As a result, the task-irrelevant features of stay/switch cues was not diagnostic of the presence of these cues in any given trial. After these modifications, we had 129 main images comprising 43 rooms (40 main rooms plus three room ‘backgrounds’ from the art stay/switch cues) each paired with multiple paintings. Likewise, each of the 43 paintings (40 main paintings plus three paintings embedded in room stay/switch cues) were paired with multiple rooms.

20 images (unique art and room combinations) were chosen as ‘base images.’ These were used to create 20 ‘base sets’ with 7 images each: a base image, a room match (an image with the same spatial layout as the base image, from a different perspective), an art match (an image with a painting by the same artist as the base image) and 4 distractors (rooms with different layouts and different artists compared to the base image). Room and art matches in one base set could be distractors in another base set. Base images were not used as distractors or matches in other base sets. An image that was an art match to the base image could not also be a room match to the base image, or vice versa. A given trial consisted of the presentation of a base image and 4 ‘search’ images (from the pool of: art match, room match, distractors, stay/switch cue). Each base set was used to generate 10 trials: 5 in the memory-guided task and 5 in the explicitly instructed task.

A nonoverlapping set of 82 images (70 images plus 12 stay/switch cue images) were used during an initial practice day (~2 days before the fMRI scan). 70 main images were separated into 10 base sets of 7 images each (a base image, an art match, a room match, and 4 distractors). As in the scan session, 12 stay/switch cue images were generated by pairing each art cue (1 stay and 1 switch) with 3 different rooms, and each room cue (1 stay and 1 switch) with 3 different paintings. However, the stay/switch cues for this practice session were distinct from those used in the fMRI scan. The purpose of this session was to give individuals practice with the task, without exposing them to the specific stimuli to be used in the scanner.

An additional nonoverlapping set of 82 images (70 main images plus 12 stay/switch cue images) were used for a practice session that took place just before the fMRI scan. The 70 main images did not overlap with either the scan day images nor the initial practice day images (~2 days before the scan). The art and room stay/switch cues for this practice session were identical to those used during the fMRI scan. However, they were paired with rooms and paintings that were part of the 70 practice-specific images, which did not overlap with those used in the fMRI scan or the initial practice day. As in the other sessions, the art stay/switch cues were each paired with 3 rooms, and the room stay/switch cues were each paired with 3 paintings, making 12 stay/switch cue images in total.

Stimuli were presented using the Psychophysics Toolbox for MATLAB (psychtoolbox.org). At the beginning of each explicitly instructed trial, participants received the instruction to”Press any key with left hand to start the trial’. At the beginning of each memory-guided trial, participants received the instruction to ‘Press left index for Room, left middle for Art’. This initiation screen remained visible until the participant responded. Apart from the initiation screen, the rest of the trial was identical for the explicitly instructed and memory-guided tasks.

After a key was pressed on the initiation screen in the explicitly instructed task, the attentional cue (‘ART’ or ‘ROOM’) was randomly assigned. In the memory-guided task, participants were instructed to select their attentional state based on the stay/switch cue in the preceding trial. This is similar to task-switching studies in which a cue (often an abstract one) signals when participants should switch to doing a different task (Chiu and Yantis, 2009; Monsell, 2003). For example, if the attentional state on the previous trial was ‘art’, and there was an art ‘switch’ cue, then the attentional state on the current trial should be ‘room’ (art stay/switch cues only appeared on trials where art was attended; room stay/switch cues only appeared on trials where rooms were attended). If the participant mistakenly selected ‘art’, then the trial proceeded with an art attentional state. One-third of the trials did not contain a stay or switch cue. In the memory-guided task, following these ‘no-cue’ trials, and also on the first trial of each run, participants were free to choose whichever attentional state they wanted, but they were instructed to choose art and room approximately equally often. These no-cue trials were included in the design to test additional hypotheses beyond the focus of the present paper. Following these ‘no-cue’ trials, participants on average chose room (M = 16.828, 95% CI [16.299, 17.356]) more often than art (M = 14.655, 95% CI [14.144, 15.166]), t(28) = 5.90, p<0.00001, d = 1.10, 95% CI [1.418, 2.927]. However, this imbalance was only a few trials per participant (median = 3, min = 0, max = 7). Nevertheless, art and room trials were equally weighted in all analyses, so this slight difference could not account for any observed effects.

Following the initiation button press, participants were presented with the attentional cue, (‘ART’ or ‘ROOM’, centered at fixation), which remained on the screen for either 1.5 s, 2 s, or 2.5 s, randomized across trials. After the attentional cue, a base image was presented for 2 s. Then, four search images, centered at fixation, were presented for 1.25 s each, separated by 0.1 s inter-stimulus intervals. The ‘ART?’ or ‘ROOM?’ probe was then presented 0.1 s after the offset of the last search image, for a maximum of 2 s (less if the participant responded within that time). Participants indicated if there was a match present or absent by pressing the button box with the right-hand index or middle finger, respectively.

When the probe was ‘ART?”, participants' goal was to indicate if any of the paintings in the search images could have been painted by the same artist who painted the painting in the base image. For ‘ROOM?’ probes, participants’ goal was to indicate if any of the room layouts in the search images was the same as that of the base image, but from a different perspective. 80% of trials were ‘valid’ trials, in which the attentional cue at the beginning of the trial matched the probe at the end. 20% of trials were ‘invalid’ trials, in which the attentional cue at the beginning of the trial did not match the probe at the end. This allowed us to ensure that attention was effectively engaged by the cue at the beginning of the trial (Figure 2).

Trials were separated by a blank inter-trial interval (ITI) of variable length. For each experimental run, the same set of 25 ITIs (truncated exponential, lambda = 1.5, mean = 6.66 s, T = 9 s) was used in a random order. At the end of each run, the percentage of correct responses was presented. In memory-guided runs, the accuracy of selecting the correct attentional state (based on the stay/switch cues) was also presented.

The same trial structure was used for the practice sessions except that the ITI was either 2 s or 2.5 s randomly determined on each trial. Furthermore, feedback on practice trials was shown after each probe (e.g., ‘Correct, there was an art match’), and for the memory-guided task after each attentional state selection (e.g., ‘Correct, there was an art switch cue on the previous trial’).

Trial order was randomized with two constraints: (i) Each of the 20 base images was shown once every 20 trials, and (ii) the same base image was not repeated across two consecutive trials. The task-relevant match (e.g., an art match on a trial with an art probe) was shown on half of the trials, and, independently, the task-irrelevant match (e.g., a room match on a trial with an art probe) was shown on half of the trials. The remaining images in the search set were distractors, chosen among the four distractor image options for a given base set. A given image was never shown twice in a trial. On two-thirds of the trials, one of the 12 stay/switch cue images replaced one of the distractor images (this was true for both the memory-guided and explicitly instructed tasks; for the explicitly instructed task, these stay/switch cues had no relevance for the attentional cue on the following trial).

Valid trials of each task (memory-guided, explicitly instructed) were distributed across the two attentional states (art; room), two task-relevant match types (match present; match absent), two task-irrelevant match types (match present; match absent), and three cue types (stay; switch; none) as equally as possible. Although perfectly equating trial numbers across conditions was not possible for a given participant, trial numbers were equated every six participants.

Participants first came in for a behavioral practice session approximately 2 days before the fMRI scan. This session involved training in both the memory-guided and explicitly instructed tasks, but with stimuli that were non-overlapping with those used in the fMRI session. Both the practice session and the fMRI session followed the same procedure (below).

On each practice session (~2 days before the fMRI scan and on the day of the fMRI scan), participants completed 3 phases of practice. First, they completed a run of 10 trials in the explicitly instructed task. This run was repeated until participants reached at least 65% accuracy on validly cued trials. Next, participants completed the stay/switch cue learning phase. Here, the stay/switch cue images and their meanings (i.e., stay or switch) were presented for four times each in shuffled order, for a minimum of 1 s. The participant had to push a button to continue to the next image. Then, the stay/switch cue images were shown again, this time without their meanings (i.e., no stay/switch label), five times each in shuffled order. Participants indicated if a given image was a stay or switch cue with a button press. Completion of this phase required responding accurately to every image 5 times in a row. Upon a single incorrect response, this test phase was terminated, and the stay/switch cue learning phase was restarted from the beginning by presenting each stay/switch cue image and its meaning for 4 times. After completing the stay/switch cue test, participants performed a run of 10 trials in the memory-guided task. This memory-guided practice session ended once participants reached, in a given run of 10 trials, a minimum of 65% accuracy for validly cued trials in the attention task and a minimum of 80% accuracy for selecting the correct attentional state based on stay/switch cues.

Participants then completed the fMRI task, for which there were 8 runs of 25 trials each. Explicitly instructed (100 trials) and memory-guided (100 trials) tasks were blocked to constitute either the first or second half of the experimental session (order counterbalanced across participants). When starting a new task, participants performed five practice trials to get used to that particular task. The practice was repeated until accuracy on the art/room attention task was at least 65%. In memory-guided runs, the practice was also repeated until accuracy in selecting the appropriate attentional state based on stay/switch cues was at least 80%. At the end of each memory-guided task run, participants were shown a reminder screen with all four stay/switch cues (two paintings, two rooms) and their meanings (i.e., stay or switch). If on a given memory-guided task run, the average accuracy of choosing the correct attentional state was less than 85%, then the stay/switch cue learning phase (mentioned in the previous paragraph) was repeated.

Our design has several important aspects. First, a room match and an art match were equally and independently likely to be present in search images, for both art trials and room trials. Thus, accurate responding required being in the correct attentional state. Second, the same stimuli were used for the art and room attentional states (except for the stay/switch cues), so that differences in brain activity for these states must reflect top-down attentional goals rather than differences in the stimuli presented. Third, stimuli were identical across the memory-guided and explicitly instructed tasks, including the stay/switch cues. However, in the explicitly instructed task, the presence of a stay or switch cue did not have any consequence for participants’ attentional states (because these states were randomly assigned on each trial). Thus, differences in brain activity between the memory-guided and explicitly instructed tasks cannot be due to the mere presence of stay/switch cues, but rather must be due to the need to use these cues to guide attention. Finally, motor demands were the same for the memory-guided and explicitly instructed tasks. Thus, the only difference between these tasks was the need to use memory to guide attention.

Preprocessing and analyses were performed using FEAT, FNIRT, and command-line functions in FSL (e.g., fslmaths). ROI (region of interest) analyses (e.g., univariate activity, pattern similarity, and multivariate connectivity) were performed using custom Matlab scripts. Data, experiment code, and analysis code are publicly available on the Open Science Framework: https://osf.io/ndf6b/.

The hippocampus ROI was anatomically defined from the Harvard-Oxford atlas in FSL (Jenkinson et al., 2012). The vmPFC ROI was based on Mackey and Petrides (2014), but we removed voxels that overlapped with the corpus callosum. The V1-2 ROI was obtained from the human visual cortex atlas provided in Wang et al. (2015). ROIs are shown in Figure 7.

Figure 7

Download asset Open asset

The first 4 volumes of each run were discarded to allow for T1 equilibration (except for one participant for whom only one extra volume, rather than 4, was collected for this reason). Brain extraction, motion correction (using the MCFLIRT motion correction tool of FSL; Jenkinson et al., 2002), high-pass filtering (cut-off = 128 s), and spatial smoothing (3 mm FWHM Gaussian kernel) were performed as preprocessing steps. Field map preprocessing was based on recommendations in the FUGUE user guide (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/FUGUE/Guide) and carried out with a custom script. First, two magnitude images were averaged and skull stripped. The average magnitude image was then used together with the phase image to generate a field map image using the fsl_prepare_fieldmap command of FSL. This field map image and the average magnitude image were included in the preprocessing step of FEAT analyses to unwarp the functional images and aid registration to anatomical space. This approach helped to reduce the distortion in anterior temporal and frontal regions. Functional images were registered to the standard MNI152 T1-weighted structural image using a non-linear warp with a resolution of 10 mm and 12 degrees of freedom.

Only valid trials were used for image period analyses, to reduce any potential BOLD signal contamination from an invalid probe (as in Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b; note, however, that our results hold when all trials are used). We included trials with both correct and incorrect responses to balance the number of trials per participant. To test if univariate activity levels were higher for memory-guided vs. explicitly instructed tasks, we modeled the data with a single-trial GLM. Each trial (25 in each run) was modeled as a 7.4 s epoch from the onset of the base image to the offset of the last search image. There were two additional regressors: a regressor for all orienting periods, modeled as the interval from the onset of the initiation screen (which remained until a key was pressed) until the offset of the attentional cue; and a regressor for all probe periods, modeled as a 2 s epoch during the probe display. All regressors were convolved with a double-gamma hemodynamic response function. Finally, the 6 directions of head motion were included as nuisance regressors. Autocorrelations in the time series were corrected with FILM prewhitening. Each run was modeled separately, resulting in eight different models per participant. For each participant, image period parameter estimates were averaged across voxels within each ROI, and the resulting values for the memory-guided and explicitly instructed tasks were compared at the group level using a paired-samples t-test.

To test if the activity enhancement for memory-guided vs. explicitly instructed tasks was correlated between the hippocampus and vmPFC across individuals, we first subtracted the average parameter estimate in the explicitly instructed task from that of the memory-guided task for each individual, separately for the hippocampus and vmPFC. Then, these memory-guided vs. explicitly instructed difference scores in the hippocampus and vmPFC were correlated across individuals using the skipped_correlation.m function (https://github.com/CPernet/robustcorrtool; Pernet et al., 2012). This function performs a robust correlation by removing bivariate outliers as determined by: (1) finding the central point in the distribution using the minimum covariance determinant (Rousseeuw and Driessen, 1999), (2) orthogonally projecting each data point onto lines that join each data point to the estimated center point, (3) identifying outliers on the projected data using the interquartile range method (Frigge et al., 1989), and (4) calculating Pearson’s correlation after removing the outlier(s). With this approach, one participant was excluded as an outlier. However, our results hold when this participant was included in the analysis (R² = 0.44, p=0.000092, 95% CI [0.39, 0.83]).

To test if multivariate patterns of activity represent online attentional goals, we conducted pattern similarity analysis on trial-specific activity patterns from the image periods. This, and all other multivariate analyses, were conducted on preprocessed fMRI data. First, preprocessed data were z-scored across time, within each voxel, separately for each run. Data were then shifted forward by 6 s (4 TRs) to account for hemodynamic lag before selecting TRs that corresponded to each image period. Specifically, TRs for which at least half of the volume acquisition time corresponded to the (shifted) image period were considered to be image period TRs. Mean activity patterns for each image period were obtained for each region of interest by averaging activity levels for each voxel across all image period TRs. These activity patterns were then reshaped into a single-dimensional vector (length = number of voxels). Pearson correlations were then calculated between all pairs of vectors (i.e., between all trials) in different runs of the same task (i.e., task-specific pattern similarity was obtained by comparing explicitly instructed runs to other runs of the same task, and memory-guided runs to other runs of the same task). Correlations between trial pairs within the same run were excluded to reduce the effects of autocorrelation on pattern similarity values (Mumford et al., 2014). Finally, correlations were averaged separately for same state trial pairs (i.e., art-art, room-room) and different state trial pairs (i.e., art-room). For statistical testing, correlations were Fisher-transformed before averaging to ensure normality. Fisher-transformed pattern similarity values for trials of the same vs. different attentional states were compared at the group level with a paired-samples t-test, separately for the memory-guided and explicitly instructed conditions.

We used only valid trials for the image period analysis — as in our previous work (Aly and Turk-Browne, 2016a; Aly and Turk-Browne, 2016b) — because neural activity on invalid trials might partly reflect the invalidly probed attentional state representation. However, we obtained the same pattern of results when both invalid and valid trials were included in the analysis.

To examine interactions between the hippocampus and early visual cortex, we calculated multivariate connectivity between these regions. In order to do this, we first obtained measures of the ‘quality’ of attentional states in each region on a trial-by-trial basis, and then correlated these attentional state quality measures between regions.

For this analysis, we used the z-scored, preprocessed data as we did for the image period pattern similarity analysis mentioned above. First, we averaged activity patterns across trials, separately for art and room attentional states, to create art and room attentional state ‘templates’. These templates indicate what brain activity in a given region generally looks like for the art vs. room attentional states. Second, we calculated Pearson correlations between these templates and activity patterns for individual trials. Importantly, the templates compared to a given trial excluded trials in the same run (e.g., for analysis of a trial in run 1, templates excluded other trials in run 1; Mumford et al., 2014). Third, for each trial’s activity pattern, we calculated a measure of multivariate attentional state ‘quality’ by subtracting its correlation with the different-state template (e.g., an art trial correlated with the room template) from its correlation with the same-state template (e.g., an art trial correlated with the art template). These steps were performed separately for the hippocampus and V1-2. Lastly, we computed multivariate connectivity between the hippocampus and V1-2 by calculating Pearson correlations for their multivariate attentional state ‘quality’ scores across all trials. For statistical testing, multivariate connectivity values were Fisher-transformed to ensure normality. The Fisher-transformed connectivity values for the explicitly instructed and memory-guided tasks were compared at the group level using a paired-samples t-test. These values were also compared to zero using a one-sample t-test for each task.

To examine whether preparatory univariate activity in the hippocampus was enhanced for memory-guided vs. explicitly instructed attention (Stokes et al., 2012), we examined BOLD activity in the hippocampus during the orienting period. To this end, we performed a single-trial GLM with 27 regressors. There were 25 orienting period regressors (one for each orienting period), modeled from the onset of the initiation screen (which remained until a key was pressed) until the offset of the attentional cue; a single regressor for all image periods, modeled as 7.4 s epochs from the onset of the base image to the offset of last search image; and a regressor for all probe periods, modeled as 2 s epochs during the probe displays. As in the image period analyses, (i) all regressors were convolved with a double-gamma hemodynamic response function, (ii) the 6 directions of head motion were included as nuisance regressors, (iii) autocorrelations in the time series were corrected with FILM prewhitening, (iv) both correct and incorrect responses were included, (v) only valid trials were used (our results hold when invalid trials are included), and (vi) each run was modeled separately. The first trial of each run was excluded from the orienting period analysis, as there was no previous trial for the attentional state decision to be based on (in the memory-guided task). For each participant, orienting period parameter estimates were averaged across voxels, and the resulting values for the memory-guided and explicitly instructed tasks were compared at the group level using a paired-samples t-test. For completeness, we also performed this analysis for vmPFC.

To test if multivariate activity patterns during the orienting period represented preparatory attentional states, activity patterns during the orienting periods were correlated with activity patterns from the image periods. As in the image period analysis, we used preprocessed and z-scored data.

Given that we were interested in the correlation between the activity patterns of these two temporally adjacent periods, we attempted to limit their autocorrelation — induced by the slow hemodynamic response — in two ways. First, we only compared orienting period activity patterns and image period activity patterns across runs, i.e., the orienting period activity patterns on run 1 were never compared to image period activity patterns in run 1 (Mumford et al., 2014). Second, to further reduce their autocorrelation, we removed boundary TRs from the analysis (i.e., the last TR of the orienting period and the first and last TR of the image period; it was not necessary to drop the first TR of the orienting period because it followed a blank inter-trial interval). The first TR of the image period was removed to reduce autocorrelation with the orienting period, which is important given that we used the correlation between these two periods as evidence for preparatory attentional states. The last TR of the image period was not included so as to remove BOLD activity due to the probe. This is particularly critical when comparing orienting period activity patterns to image period activity patterns as a marker of preparatory attentional states: because the cue component of the orienting period (‘ART’ or ‘ROOM’) overlaps perceptually with the probe (‘ART?’ or ‘ROOM’?), not dropping the last TR of the image period risks an artificial boost of orienting period/image period pattern similarity as a result of this perceptual overlap. Note that a TR was considered to be part of the image period if at least 50% of the duration of that brain volume acquisition corresponded to the image period, but this still leaves a considerable amount of time for the probe to affect brain activity during that TR. For these reasons, the last TR of the image period was dropped in order to be conservative.

Importantly, the last TR of the orienting period was removed only if it was not the only TR during which the attentional cue was presented. This ensured that orienting period activity patterns always included timepoints at which the attentional state was known to the participant. This is particularly important for the explicitly instructed condition: Otherwise, a difference between the memory-guided and explicitly instructed tasks could simply arise because participants know their attentional state in one task but not in the other. Thus, this step ensured that any differences between tasks are because of how attentional state information was obtained (from memory or an overt instruction), rather than its availability.

After dropping boundary TRs in this way, we obtained a mean activity pattern for each period (orienting or image) by averaging over the remaining TRs. Because we dropped boundary TRs in this analysis, but not in the main image period analysis, we first confirmed that the hippocampus and vmPFC still discriminate between the art and room attentional states during the image period with this new, conservative approach. Indeed, both the hippocampus (memory-guided: t(28) = 2.87, p=0.00078, d = 0.53, 95% CI [0.001, 0.007], explicitly instructed: t(28) = 6.59, p<0.0001, d = 1.22, 95% CI [0.007, 0.013]) and vmPFC (memory-guided: t(28) = 6.61, p<0.0001, d = 1.23, 95% CI [0.009, 0.017], explicitly instructed: t(28) = 6.06, p<0.0001, d = 1.13, 95% CI [0.011, 0.023]) still exhibited greater pattern similarity for trials of the same vs. different attentional states in both tasks.

Having confirmed distinct representations for the art and room attentional states with this approach, we next obtained ‘template’ activity patterns for the image periods. These template activity patterns were the average of valid art trials (for the art template) and the average of valid room trials (for the room template). The purpose of these templates was to obtain activity patterns that represent online attention to artistic styles vs. room layouts.

Activity patterns for the orienting period of each trial were then correlated with the art template and room template. Importantly, the image period templates excluded all trials in the same run as a given orienting period activity pattern (e.g., for the analysis of the orienting period in run 1, image period templates excluded trials in run 1). The correlations between the orienting period activity patterns and the image period templates were then grouped based on whether they were a match to the same state (e.g., an art trial orienting period activity pattern correlated with an art image period template) or a match to the different state (e.g., an art trial orienting period activity pattern correlated with a room image period template). This was repeated for all trials. The correlations were then averaged separately for each combination of match type (match to same template; match to different template), attentional state (art; room), and task (explicitly instructed; memory-guided).

These pattern similarity values were averaged across attentional states (art, room). This ensured that art and room trials contributed to average pattern similarity values equally. To measure preparatory attentional states, we calculated the difference in average pattern similarity for match-to-same template and match-to-different template correlations, separately for each participant. The match-to-same template and match-to-different template difference scores were then compared for the memory-guided and explicitly instructed tasks with a paired-samples t-test, after Fisher-transforming these values to ensure normality. Finally, the difference scores were compared to zero using one-sample t-tests, separately for memory-guided and explicitly instructed tasks. Values significantly above 0 indicate evidence for the upcoming attentional state.

To test whether other brain regions represent preparatory attentional states, we performed the orienting period pattern similarity analysis using a whole-brain searchlight approach, via the Simitar toolbox (Pereira and Botvinick, 2013). This analysis was identical to the main orienting period ROI analysis (Figure 1—figure supplement 1) except that pattern similarity was calculated for all possible 27-voxel cubes (3×3×3 voxels) throughout the brain. The result (i.e., orienting period to image period pattern similarity) for each cube was assigned to the center voxel. This analysis was conducted separately for each participant, and group-level statistics were then performed with the randomise function in FSL. Specifically, we performed a non-parametric one-sample t-test that used 10,000 permutations to generate a null distribution. Voxel-based thresholding was applied, corrected for multiple comparisons using the family-wise error rate correction (p<0.05).

• Mariam Aly

• Mariam Aly