Neural representations of naturalistic events are updated as our understanding of the past changes

  1. Asieh Zadbood
  2. Samuel Nastase
  3. Janice Chen
  4. Kenneth A Norman
  5. Uri Hasson

  • Curated by eLife

    eLife logo

    Evaluation Summary:

    This is a well-designed study that will be of interest to cognitive neuroscientists studying event perception and memory, particularly those interested in naturalistic paradigms. The main contribution is in growing our understanding of how interpretational differences of events are reflected in differences in neural representations of those events. While the presented results are convincing, it remains somewhat unclear what processes drive the observed effects, and thus what the role of the implicated brain regions is in memory updating.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #3 agreed to share their name with the authors.)

Reviewed by

Read the full article See related articles

Listed in

Log in to save this article

Abstract

The brain actively reshapes our understanding of past events in light of new incoming information. In the current study, we ask how the brain supports this updating process during the encoding and recall of naturalistic stimuli. One group of participants watched a movie (‘The Sixth Sense’) with a cinematic ‘twist’ at the end that dramatically changed the interpretation of previous events. Next, participants were asked to verbally recall the movie events, taking into account the new ‘twist’ information. Most participants updated their recall to incorporate the twist. Two additional groups recalled the movie without having to update their memories during recall: one group never saw the twist; another group was exposed to the twist prior to the beginning of the movie, and thus the twist information was incorporated both during encoding and recall. We found that providing participants with information about the twist beforehand altered neural response patterns during movie-viewing in the default mode network (DMN). Moreover, presenting participants with the twist at the end of the movie changed the neural representation of the previously-encoded information during recall in a subset of DMN regions. Further evidence for this transformation was obtained by comparing the neural activation patterns during encoding and recall and correlating them with behavioral signatures of memory updating. Our results demonstrate that neural representations of past events encoded in the DMN are dynamically integrated with new information that reshapes our understanding in natural contexts.

Article activity feed

  1. Version published to 10.7554/elife.79045 on eLife
  2. eLife

    Author Response

    Reviewer #1 (Public Review):

    This fMRI study investigated how memories are updated after reinterpreting past events. Participants watched a movie and subsequently recalled individual scenes from that movie. Importantly, the movie ends with a twist that changes the interpretation of earlier scenes in the movie. One group of participants watched the movie with the twist at the end, one group did not get to see the twist, and a third group was already informed about this twist before watching the movie. Analyses compared the similarity of activity patterns to (encoded or recalled) events across participants within regions of the default mode network (DMN). The design allowed for multiple relevant comparisons, confirming the prediction that activity patterns in DMN regions reflect the (re)interpretation of the movie (during movie viewing and/or during recall).

    The study is well-designed and executed. The inclusion of multiple analyses involving distinct comparisons strengthens the evidence for the role of the DMN in memory updating.

    The following points may be relevant to consider:

    1. The cross-participant pattern analysis method used here is not standard, with such analyses typically done within participants (or across participants, but after aligning representational spaces). Considering individual variability in functional organization, the method is likely only sensitive to coarse-scale patterns (e.g., anterior vs posterior parts of an ROI). This is not necessarily a weakness but is relevant when interpreting the results.

    We agree with the reviewer that functional misalignment might have played against us here. We designed this study as a natural successor of our previous work in which we captured reliable and multimodal scene-specific cross-participant pattern similarity during encoding and recall in standard space. In this revised version, we provide further evidence on how scene content is captured and influences our results. Nonetheless, we agree with your comment and add the following section to the discussion to encourage considering this point while interpreting the results.

    "Moreover, our current method relies on averaging spatially-coarse activity patterns across subjects (and time points within an event). Future extensions of this work may benefit from using functional alignment methods (Haxby et al 2020, Chen et al 2015) to capture more fine-grained event representations which are shared across participants."

    1. Unlike previous work, analyses are not testing for scene-specific information. Rather, each scene is treated separately to establish between-group differences, and results are averaged across scenes. This raises the question of whether the patterns reflect scene-specific information or generic group differences. For example, knowing the twist may increase overall engagement, both when viewing the movie (spoiled group) and when recalling it (spoiled group + twist group). The DMN may be particularly sensitive to such differences in overall engagement.

    You have brought up great points. We addressed them in two ways: (1) We ran a univariate analysis in each DMN ROI to look at the role of overall regional-average response magnitude in our results. We did not observe a significant effect of group or an interaction between group and condition. (2) We ran a scene-specificity analysis in a new Results section entitled “The role of scene content” (Figure 4). This section is focused on comparing interaction index (Figure 2C), as an indicator of memory updating, under different manipulations. Interaction index reflects the reversal of neural similarity during encoding and recall. Our results suggest that we don’t see the same effects if we shuffle the scene labels and recompute the pattern similarity analyses. Please see added text and figures below:

    "To test whether our reported results were mainly driven by the similarities and differences in multivariate spatial patterns of neural representations, as opposed to by univariate regional-average response magnitudes, we ran a univariate analysis in each ROI. This analysis revealed no significant effect of group (“spoiled”, “twist”, “no-twist”) or interaction between group and condition (movie, recall) (Table 1, see Methods for details).

    Next, to determine whether scene-specific neural event representations—as opposed to coarser differences in general mental state across all scenes with similar interpretations—drive our observed pISC differences, we shuffled the labels of critical scenes within each group before calculating and comparing pISC across groups. By repeating this procedure 1000 times and recalculating the interaction index at each iteration, we constructed a null distribution of interaction indices for shuffled critical scenes (light magenta distributions in Figure 4B). In 12 out of 24 DMN regions, interaction indices were statistically significant based on the shuffled-scene distribution (p < .025, FDR controlled at q < .05). All of these 12 regions were among the ROIs that showed meaningful effects in our original analysis (Figure 2C). Regions with significant scene-specific interaction effects are marked as blue dots with black borders in Figure 4B. Overall, the findings from this analysis confirm that our results are driven by changes to scene-specific representations."

    1. The study does not reveal what the DMN represents about the movie, such that its activity changes after knowing the twist. The Discussion briefly mentions that it may reflect the state of the observer, related to the belief about the identity of the doctor. This suggests a link to the theory of mind/mentalizing, but this is not made explicit. Alternatively, the DMN may be involved in the conflict (or switching) between the two interpretations.

    Great points. We added to the discussion about the role of mentalizing network and in the particular temporo-parietal cortex. About your last point, we think our whole brain findings outside DMN (ACC and dlPFC) might relate to that point. We discussed these further in the paper.

    "We performed two targeted analyses to look for evidence of memory updating across encoding and recall: the interaction analysis (Figure 2C) and the encoding-recall analysis (Figure 3). We hypothesized that a shift in direction of pISC difference would occur when neural representations during recall in the “twist” group start to reflect the Ghost interpretation. The interaction analysis probed this shift indirectly by taking into account the effects of both encoding-encoding and recall-recall analyses. Unlike the interaction analysis, in the encoding-recall analysis, we directly compared neural event representations during encoding and recall. Interestingly, all regions exhibiting an effect across the two encoding-recall analyses, excluding left anterior temporal cortex, were present in the interaction results. Among these regions, the left angular gyrus/TPJ exhibited an effect across all three analyses. As a core hub in the mentalizing network, temporo-parietal cortex has been implicated in theory of mind through perspective-taking, rationalizing the mental state of someone else, and modeling the attentional state of others (Frith and Frith 2006, Guterstam et. al 2021, Saxe and Kanwisher 2003). The motivations behind some actions of the main character in the movie heavily depend on whether the viewer perceives them as a Doctor or a Ghost, and participants may focus on this during both encoding and recall. We speculate that neural event representations in AG/TPJ in the current experiment may be related to mentalizing about the main character’s actions. Under this interpretation, the updated event representations during recall following the twist would be more closely aligned to the “spoiled” encoding representations, as a consequence of memory updating in the “twist” group.

    In our whole brain analysis, these regions did not have significant interaction effects, which suggests that the effects were isolated to encoding. In the whole-brain analysis, we also observed a significant encoding-encoding and interaction effects in anterior cingulate cortex, as well as recall-recall and interaction effects in dlPFC. These results suggest that both the "spoiled" manipulation and the "twist" may recruit top-down control and conflict monitoring processes during naturalistic viewing and recall."

    1. The design has many naturalistic aspects, but it is also different from real life in that the critical twist involves a ghost. Furthermore, all results are based on one movie with a specific plot twist. It is thus not clear whether similar results would be obtained with other and more naturalistic plot twists.

    We added this as a limitation of the study.

    "Our findings provide further insight into the functional role of the DMN. However, these results have been obtained using only one movie. While naturalistic paradigms better capture the complexity of real life and provide greater ecological generalizability than highly-controlled experimental stimuli and tasks (Nastase et al., 2020), they are still limited by the properties of the particular naturalistic stimulus used. For example, this movie—including the twist itself—hinges on suspension of disbelief about the existence of ghosts. Future work is needed to extend our findings about updating event memories to a broader class of naturalistic stimuli: for example, movies with different kinds of (non-supernatural) plot twists, spoken stories with twist endings, or using autobiographical real-life situations where new information (e.g. discovering a longtime friend has lied about something important) triggers re-evaluation of the past (e.g. reinterpreting their friend’s previous actions)."

    1. Only 7 scenes (out of 18) were included in the analysis. It is not clear if/how the results depend on the selection of these 7 scenes.

    Thank you for bringing this up. These scenes were pre-selected for the analyses, as they are the only scenes that are rated high by our independent raters (not study participants) on “twist influence”, meaning that knowing the twist could dramatically change their interpretation. So, we had a priori reasons to hypothesize that the effect will be strong in these scenes. To address your point, we report results by including all 18 scenes in a new Results section entitled “The role of scene content” and in Figure 4A. While the effect was weaker for all scenes it was still apparent in this conservative analysis. As expected, however, including 7 critical scenes produces stronger results than including all scenes or the uncritical scenes (all minus critical scenes). Please see the “The role of scene content” in Results and in Figure 4 for more detailed information.

    "The role of scene content In the prior analyses, we focused on “critical scenes”, selected based on ratings from four raters who quantified the influence of the twist on the interpretation of each scene (see Methods). An independent post-experiment analysis of the verbal recall behavior of the fMRI participants yielded “twist scores” that were also highest for these scenes; that is, the expected and perceived effect of twist information on recall behavior were found to match. In our next analysis, we asked whether the neural event representations reflect these differences in the twist-related content of the scenes. In other words, are the “critical scenes” with highly twist-dependent interpretations truly critical for our observed effects?

    To answer this question, we re-ran our main encoding-encoding and recall-recall pISC analysis in each DMN ROI (Figure 2-3). We calculated interaction indices (Figure 2C) first by including all scenes, and second by including only the 11 non-critical scenes. To better compare the effect of including different subsets of scenes to our original results, in Figure 4 we show the results in 15 ROIs that exhibited meaningful effects in our main analyses (Figure 2C). Figure 4A demonstrates that “critical scenes” yielded higher interaction indices compared to all scenes or non-critical scenes across all ROIs. The interaction score across all DMN ROIs was significantly higher in “critical scenes” than all scenes (t(23) = 7.19, p = 2.53 x 10-7) and non-critical scenes (t(23) = 7.3, p = 1.95 x 10-7). These results show that critical scenes are indeed responsible for the observed pISC differences across groups."

    Reviewer #2 (Public Review):

    In this manuscript titled "Here's the twist: How the brain updates the representations of naturalistic events as our understanding of the past changes", the authors reported a study that examined how new information (manipulated as a twist at the end of a movie) changes the neural representations in the default mode network (DMN) during the recall of prior knowledge. Three groups of participants were compared - one group experienced the twist at the end, one group never experienced the twist, and one group received a spoiler at the beginning. At retrieval, participants received snippets of 18 scenes of the movie as cues and were asked to freely describe the events of each scene and to provide the most accurate interpretation of the scene, given the information they gathered throughout watching.

    All three groups were highly accurate in the recall of content. The groups that experienced the twist at the end as well as at the beginning as a spoiler showed a higher twist score (the extent to which twist information was incorporated into the recall), while seemingly also keeping the interpretation without the twist ("Doctor representation") intact. Neurally, several regions in the DMN showed significant interaction effects in their neural similarity patterns (based on intersubject pattern correlation), indicating a change in interpretation between encoding and recall in the twist group uniquely, presumably reflecting memory updating.

    Several points that I think should be addressed to strengthen the manuscript:

    1. The results from encoding-retrieval similarity analysis (particularly the one depicted in Figure 3B) don't match the results from encoding/retrieval interaction (particularly those shown in Figure 2C). While they were certainly based on different comparisons, I would think that both analyses were set up to test for memory updating. Can the authors comment on this divergence in results?

    Thank you for your comment. Except for one ROI, the other two regions in Figure 2C are present in the interaction analysis. The ROI at the frontal pole might be hard to see from this angle but in fact it holds a high effect size in interaction analysis. So we do not see a big divergence between these two results. But taking into account the recall-recall results, we agree that there seems to be inhomogeneity. We discussed these further in the discussion.

    "We performed two targeted analyses to look for evidence of memory updating across encoding and recall: the interaction analysis (Figure 2C) and the encoding-recall analysis (Figure 3). We hypothesized that a shift in direction of pISC difference would occur when neural representations during recall in the “twist” group start to reflect the Ghost interpretation. The interaction analysis probed this shift indirectly by taking into account the effects of both encoding-encoding and recall-recall analyses. Unlike the interaction analysis, in the encoding-recall analysis, we directly compared neural event representations during encoding and recall. Interestingly, all regions exhibiting an effect across the two encoding-recall analyses, excluding left anterior temporal cortex, were present in the interaction results. Among these regions, the left angular gyrus/TPJ exhibited an effect across all three analyses. As a core hub in the mentalizing network, temporo-parietal cortex has been implicated in theory of mind through perspective-taking, rationalizing the mental state of someone else, and modeling the attentional state of others (Frith and Frith 2006, Guterstam et. al 2021, Saxe and Kanwisher 2003). The motivations behind some actions of the main character in the movie heavily depend on whether the viewer perceives them as a Doctor or a Ghost, and participants may focus on this during both encoding and recall. We speculate that neural event representations in AG/TPJ in the current experiment may be related to mentalizing about the main character’s actions. Under this interpretation, the updated event representations during recall following the twist would be more closely aligned to the “spoiled” encoding representations, as a consequence of memory updating in the “twist” group.

    Our findings are consistent with the view that DMN synthesizes incoming information with one’s prior beliefs and memories (Yeshurun et al 2021). We add to this framework by providing evidence for the involvement of DMN regions in updating prior beliefs in light of new knowledge. Across our different encoding and recall analyses, we observe memory updating effects in a varied subset of DMN regions that do not cleanly map onto a specific subsystem of DMN (Robin and Moscovitch 2017, Ranganath and Ritchey 2012, Ritchey and Cooper 2020). Rather than being divergent, these results might be reflecting inherent differences between the processes of encoding and recall of naturalistic events. It has been proposed that neural representations corresponding to encoding of events are systematically transformed during recall of those events (Chen et al 2017, Favila et al 2020, Musz and Chen 2022). While we provide evidence for reinstatement of memories in DMN, our findings also support a transformation of neural representation during recall, as encoding-recall results were weaker in some areas than recall-recall findings. This transformation could affect how different regions and sub-systems of DMN represent memories, and suggests that the concerted activity of multiple subsystems and neural mechanisms might be at play during encoding, recall and successful updating of naturalistic event memories."

    1. The recall task was self-paced. Can reaction time information be provided on how long participants needed to recall? Did this differ across groups? Presumably in the twist group and spoiled group participants might have needed a longer time to incorporate both the original and twist interpretation.

    This is an interesting idea. Unfortunately, we could not measure this accurately because our recall cues were snippets from the beginning of each scene with different length (selected based on content). And updating could begin from the beginning of those snippets (but we wouldn’t know when). We will consider this point in the future related designs.

    How was the length difference across events taken into consideration in the beta estimates?

    They were used as event durations in the GLM model.

    Also, is there an order effect, such that one type of interpretation tended to be recalled first?

    This is hard to measure as this only occurs in a subset of scenes. But we assume it happens in other people’s brains as well

    This is indeed hard to measure as you mentioned. We will provide the transcripts when sharing the data and hopefully this will facilitate future text-analysis work on this dataset to answer interesting questions like this.

    1. The correlation analysis between neural pattern change and behavioral twist score is based on a small sample size and does not seem to be well suited to test the postulation of the authors, namely that some participants may hold both interpretations in their memory. Interestingly, the twist score of the spoiled group was similar to the twist group, indicating participants in this group might have held both interpretations as well. Could this observation be leveraged, for example by combining both groups (hence better powered with larger sample size), in order to relate individual differences in neural similarity patterns and behavioral tendency to hold both interpretations?

    Even though both groups showed signs of holding both interpretations in mind, the process happening in their brain during the recall is different. In particular, we do not expect to see any updating effect in the spoiled group. So it wouldn’t seem accurate to combine these groups to test the effect of incomplete updating.

    1. Several regions within the DMN were significant across the analysis steps, specifically the angular gyrus, middle temporal cortex, and medial PFC. Can the authors provide more insights on how these widely distributed regions may act together to enable memory updating? The discussion on the main findings is largely at a rather superficial level about DMN, or focuses specifically on vmPFC, but neglects the distributed regions that presumably function interactively

    Thanks for bringing this up. We added text to discussion to respond to this very valid point. Please see the added text in our response to your first point. One more snippet added to the discussion about this:

    "In addition to mPFC, right precuneus and parts of temporal cortex exhibited significantly higher pattern similarity in the “twist” and “spoiled” groups who recalled the movie with the same interpretation. Precuneus is a core region in the posterior medial network, which is hypothesized to be involved in constructing and applying situation models (Ranganath and Ritchey 2012). Our findings support a role for precuneus in deploying interpretation-specific situation models when retrieving event memories. In particular, we suggest that the posterior medial network may encode a shift in the situation model of the “twist” group in order to accommodate the new Ghost interpretation.

    We performed two targeted analyses to look for evidence of memory updating across encoding and recall: the interaction analysis (Figure 2C) and the encoding-recall analysis (Figure 3). We hypothesized that a shift in direction of pISC difference would occur when neural representations during recall in the “twist” group start to reflect the Ghost interpretation. The interaction analysis probed this shift indirectly by taking into account the effects of both encoding-encoding and recall-recall analyses. Unlike the interaction analysis, in the encoding-recall analysis, we directly compared neural event representations during encoding and recall. Interestingly, all regions exhibiting an effect across the two encoding-recall analyses, excluding left anterior temporal cortex, were present in the interaction results. Among these regions, the left angular gyrus/TPJ exhibited an effect across all three analyses. As a core hub in the mentalizing network, temporo-parietal cortex has been implicated in theory of mind through perspective-taking, rationalizing the mental state of someone else, and modeling the attentional state of others (Frith and Frith 2006, Guterstam et. al 2021, Saxe and Kanwisher 2003). The motivations behind some actions of the main character in the movie heavily depend on whether the viewer perceives them as a Doctor or a Ghost, and participants may focus on this during both encoding and recall. We speculate that neural event representations in AG/TPJ in the current experiment may be related to mentalizing about the main character’s actions. Under this interpretation, the updated event representations during recall following the twist would be more closely aligned to the “spoiled” encoding representations, as a consequence of memory updating in the “twist” group.

    Our findings are consistent with the view that DMN synthesizes incoming information with one’s prior beliefs and memories (Yeshurun et al 2021). We add to this framework by providing evidence for the involvement of DMN regions in updating prior beliefs in light of new knowledge. Across our different encoding and recall analyses, we observe memory updating effects in a varied subset of DMN regions that do not cleanly map onto a specific subsystem of DMN (Robin and Moscovitch 2017, Ranganath and Ritchey 2012, Ritchey and Cooper 2020). Rather than being divergent, these results might be reflecting inherent differences between the processes of encoding and recall of naturalistic events. It has been proposed that neural representations corresponding to encoding of events are systematically transformed during recall of those events (Chen et al 2017, Favila et al 2020, Musz and Chen 2022). While we provide evidence for reinstatement of memories in DMN, our findings also support a transformation of neural representation during recall, as encoding-recall results were weaker in some areas than recall-recall findings. This transformation could affect how different regions and sub-systems of DMN represent memories, and suggests that the concerted activity of multiple subsystems and neural mechanisms might be at play during encoding, recall and successful updating of naturalistic event memories."

    Reviewer #3 (Public Review):

    Zadbood and colleagues investigated the way key information used to update interpretations of events alter patterns of activity in the brain. This was cleverly done by the use of "The Sixth Sense," a film featuring a famous "twist ending," which fundamentally alters the way the events in the film are understood. Participants were assigned to three groups: (1) a Spoiled group, in which the twist was revealed at the outset, (2) a Twist group, who experienced the film as normal, and (3) a No-Twist group, in which the twist was removed. Participants were scanned while watching the movie and while performing cued recall of specific scenes. Verbal recall was scored based on recall success, and evidence for descriptive bias toward two ways of understanding the events (specifically, whether a particular character was or was not a ghost). Importantly, this allowed the authors to show that the Twist group updated their interpretation. The authors focused on regions of the Default Mode Network (DMN) based on prior studies showing responsiveness to naturalistic memory paradigms in these areas and analyzed the fMRI data using intersubject pattern similarity analysis. Regions of the DMN carried patterns indicative of story interpretation. That is, encoding similarity was greater between the Twist and No-Twist groups than in the Spoiled group, and retrieval similarity was greater between the Twist and Spoiled groups than in the No-Twist group. The Spoiled group also showed greater pattern similarity with the Twist group's recall than the No-Twist group's recall. The authors also report a weaker effect of greater pattern similarity between the Spoiled group's encoding and the Twist group's recall than between the Twist group's own encoding and recall. Together, the data all converge on the point that one's interpretation of an event is an important determinant of the way it is represented in the brain.

    This is a really nice experiment, with straightforward predictions and analyses that support the claims being made. The results build directly on a prior study by this research group showing how interpretational differences in a narrative drive distinct neural representations (Yeshurun et al., 2017), but extend an understanding of how these interpretational differences might work retrospectively. I do not have any serious concerns or problems with the manuscript, the data, or the analyses. However I have a few points to raise that, if addressed, would make for a stronger paper in my opinion.

    1. My most substantive comment is that I did not find the interpretive framework to be very clear with respect to the brain regions involved. The basic effects the authors report strongly support their claims, but the particular contributions to the field might be stronger if the interpretations could be made more strongly or more specifically. In other words: the DMN is involved in updating interpretations, but how should we now think about the role of the DMN and its constituent regions as a result of this study? There are a number of ideas briefly presented about what the DMN might be doing, but it just did not feel very coherent at times. I will break this down into a few more specific points:

    While many of us would agree that the DMN is likely to be involved in the phenomena at hand, I did not find that the paper communicated the logic for singularly focusing on this subset of regions very compellingly. The authors note a few studies whose main results are found in DMN regions, but I think that this could stand to be unpacked in a more theoretically interesting way in the Introduction.

    Relatedly, I found the summary/description of regional effects in the Discussion to be a bit unsatisfying. The various pattern similarity comparisons yielded results that were actually quite nonoverlapping among DMN regions, which was not really unpacked. To be clear, it is not a 'problem' that the regional effects varied from comparison to comparison, but I do think that a more theoretical exploration of what this could mean would strengthen the paper. To the authors' credit, they describe mPFC effects through the lens of schemas, but this stands in contrast to many other regions which do not receive much consideration.

    Finally, although there is evidence that regions of the DMN act in a coordinated way under some circumstances, there is also ample evidence for distinct regional contributions to cognitive processes, memory being just one of them (e.g., Cooper & Ritchey, 2020; Robin & Moscovitch, 2017; Ranganath & Ritchey, 2012). The authors themselves introduce the idea of temporal receptive windows in a cortical hierarchy, and while DMN regions do appear to show slower temporal drift than sensory areas, those studies show regional differences in pattern stability across time even within DMN regions. Simply put, it is worth considering whether it is ideal to treat the DMN as a singular unit.

    Thank you for your helpful comments. We added text to the introduction and discussion to address your point:

    "Introduction:

    The brain’s default mode network (DMN)—comprising the posterior medial cortex, medial prefrontal cortex, temporoparietal junction, and parts of anterior temporal cortex—was originally described as an intrinsic or “task-negative” network, activated when participants are not engaged with external stimuli (Raichle et al. 2001, Buckner et al 2008). This observation led to a large body of work showing that the DMN is an important hub for supporting internally driven tasks such as memory retrieval, imagination, future planning, theory of mind, and creating and updating situation models (Svoboda et al. 2006; Addis et al. 2007; Hassabis and Maguire 2007, 2009; Schacter et al. 2007; Szpunar et al. 2007; Spreng et al. 2009, Koster-Hale & Saxe, 2013 2013, Ranganath and Ritchey 2012). However, it is not fully understood how this network contributes to these varying functions, and in particular—the focus of the present study—memory processes. Activation of this network during “offline” periods has been proposed to play a role in the consolidation of memories through replay (Kaefer et al 2022). Interestingly, prior work has also shown that the DMN is reliably engaged during “online” processing (encoding) of continuous rich dynamic stimuli such as movies and audio stories (Stephens et al 2013, Hasson et al 2008). Regions in this network have been shown to have long “temporal receptive windows” (Hasson et al 2008; Lerner et al., 2011; Chang et al., 2022), meaning that they integrate and retain high-level information that accumulates over the course of extended timescales (e.g. scenes in movies, paragraphs in text) to support comprehension. This combination of processing characteristics suggests that the DMN integrates past and new knowledge, as regions in this network have access to incoming sensory input, recent active memories, and remote long-term memories or semantic knowledge (Yeshurun et al 2021, Hasson et al 2015). These integration processes feature in many of the “constructive” processes attributed to DMN such as imagination, future planning, mentalizing, and updating situation models (Schacter and Addis 2007, Ranganath and Ritchey 2012). Notably, constructive processes are highly relevant to real-world memory updating, which involves selecting and combining the relevant parts of old and new memories. Recent work has shown that neural patterns during encoding and recall of naturalistic stimuli (movies) are reliably similar across participants in this network (Chen et al. 2017; Oedekoven et al., 2017; Zadbood et al., 2017; see Bird 2020 for a review of recent naturalistic studies on memory), and the DMN displays distinct neural activity when listening to the same story with different perspectives (Yeshurun et al 2017). Building on this foundation of prior work on the DMN, we asked whether we could find neural evidence for the retroactive influence of new knowledge on past memories."

    "Discussion :

    In addition to mPFC, right precuneus and parts of temporal cortex exhibited significantly higher pattern similarity in the “twist” and “spoiled” groups who recalled the movie with the same interpretation. Precuneus is a core region in the posterior medial network, which is hypothesized to be involved in constructing and applying situation models (Ranganath and Ritchey 2012). Our findings support a role for precuneus in deploying interpretation-specific situation models when retrieving event memories. In particular, we suggest that the posterior medial network may encode a shift in the situation model of the “twist” group in order to accommodate the new Ghost interpretation.

    We performed two targeted analyses to look for evidence of memory updating across encoding and recall: the interaction analysis (Figure 2C) and the encoding-recall analysis (Figure 3). We hypothesized that a shift in direction of pISC difference would occur when neural representations during recall in the “twist” group start to reflect the Ghost interpretation. The interaction analysis probed this shift indirectly by taking into account the effects of both encoding-encoding and recall-recall analyses. Unlike the interaction analysis, in the encoding-recall analysis, we directly compared neural event representations during encoding and recall. Interestingly, all regions exhibiting an effect across the two encoding-recall analyses, excluding left anterior temporal cortex, were present in the interaction results. Among these regions, the left angular gyrus/TPJ exhibited an effect across all three analyses. As a core hub in the mentalizing network, temporo-parietal cortex has been implicated in theory of mind through perspective-taking, rationalizing the mental state of someone else, and modeling the attentional state of others (Frith and Frith 2006, Guterstam et. al 2021, Saxe and Kanwisher 2003). The motivations behind some actions of the main character in the movie heavily depend on whether the viewer perceives them as a Doctor or a Ghost, and participants may focus on this during both encoding and recall. We speculate that neural event representations in AG/TPJ in the current experiment may be related to mentalizing about the main character’s actions. Under this interpretation, the updated event representations during recall following the twist would be more closely aligned to the “spoiled” encoding representations, as a consequence of memory updating in the “twist” group.

    Our findings are consistent with the view that DMN synthesizes incoming information with one’s prior beliefs and memories (Yeshurun et al 2021). We add to this framework by providing evidence for the involvement of DMN regions in updating prior beliefs in light of new knowledge. Across our different encoding and recall analyses, we observe memory updating effects in a varied subset of DMN regions that do not cleanly map onto a specific subsystem of DMN (Robin and Moscovitch 2017, Ranganath and Ritchey 2012, Ritchey and Cooper 2020). Rather than being divergent, these results might be reflecting inherent differences between the processes of encoding and recall of naturalistic events. It has been proposed that neural representations corresponding to encoding of events are systematically transformed during recall of those events (Chen et al 2017, Favila et al 2020, Musz and Chen 2022). While we provide evidence for reinstatement of memories in DMN, our findings also support a transformation of neural representation during recall, as encoding-recall results were weaker in some areas than recall-recall findings. This transformation could affect how different regions and sub-systems of DMN represent memories, and suggests that the concerted activity of multiple subsystems and neural mechanisms might be at play during encoding, recall and successful updating of naturalistic event memories."

    1. I think that some direct comparison to regions outside the DMN would speak to whether the DMN is truly unique in carrying the key representations being discussed here. I was reluctant to suggest this because I think that the authors are justified in expecting that DMN regions would show the effects in question. However, there really is no "null" comparison here wherein a set of regions not expected to show these effects (e.g., a somatosensory network, or the frontoparietal network) in fact do not show them. There are not really controls or key differences being hypothesized across different conditions or regions. Rather, we have a set of regions that may or may not show pattern similarity differences to varying degrees, which feels very exploratory. The inclusion of some principled control comparisons, etc. would bolster these findings. The authors do include a whole-brain analysis in Supplementary Figure 1, which indeed produced many DMN regions. However, notably, regions outside the DMN such as the primary visual cortex and mid-cingulate cortex appear to show significant effects (which, based on the color bar, might actually be stronger than effects seen in the DMN). Given the specificity of the language in the paper in terms of the DMN, I think that some direct regional or network-level comparison is needed.

    In the original submission, we included additional analyses for visual and somatosensory networks, which we hypothesized would serve as control networks. Following your comment, in the revision, we added a separate section (included below) more thoroughly examining these analyses. We also added text to the results and discussion to explain our interpretation of these findings.

    "Changes in neural representations beyond DMN We focused our core analyses on regions of the default mode network. Prior work has shown that multimodal neural representations of naturalistic events (e.g. movie scenes) are similar across encoding (movie-watching or story-listening) and verbal recall of the same events in the DMN (Chen et al., 2017; Zadbood et al., 2017). Therefore, in the current work we hypothesized that retrospective changes in the neural representations of events as the narrative interpretation shifts would be observed in the DMN. We did not, for example, expect to observe such effects in lower-level sensory regions, where neural activity differs dramatically for movie-viewing and verbal recall. To be thorough, we ran the same set of analyses we performed in the DMN (Figure 2-3) in regions of the visual and somatomotor networks extracted from the same atlas parcellation (Schaefer et al., 2018). Our results revealed larger overall differences in DMN than in visual and somatosensory networks for the key comparisons discussed previously (Figure S2). In particular, the only regions showing significant differences in pISC in recall-recall and encoding-recall comparisons (p < 0.01, uncorrected) were located in the DMN. We did not observe a notable difference between DMN and the two other networks when comparing recall “twist” to movie “spoiled” and recall “twist” to movie “twist” (RG – MG > RG – MD) which is consistent with the weak effect in the original comparison (Figure 3B). In the encoding-encoding comparison, several ROIs from the visual and somatomotor networks showed relatively strong effects as well (see Discussion).

    In addition, we qualitatively reproduced our results by performing an ROI-based whole brain analysis (Figure S3, p < 0.01 uncorrected). This analysis confirmed the importance of DMN regions for updating neural event representations. However, strong differences in pISC in the hypothesized direction were also observed in a handful of other non-DMN regions, including ROIs partly overlapping with anterior cingulate cortex and dorsolateral prefrontal cortex (see Discussion)."

    "Discussion: While our main goal in this paper was to examine how neural representations of naturalistic events change in the DMN, we also examined visual and somatosensory networks. Aside from the encoding-encoding analysis in which some visual and somatosensory regions showed stronger similarity between two groups with the same interpretation of the movie, we did not find any regions with significant effects in these two networks in the other analyses. Unlike the recall phase where each participant has their unique utterance with their own choice of words and concepts to describe the movie, the encoding (move-watching) stimulus is identical across all groups. Therefore, the effects observed during encoding-encoding analysis in sensory regions could reflect similarity in perception of the movie guided by similar attentional state while watching scenes with the same interpretation (e.g. similarity in gaze location, paying attention to certain dialogues, or small body movements while watching the movie with the same Doctor or Ghost interpretations). In our whole brain analysis, these regions did not have significant interaction effects, which suggests that the effects were isolated to encoding. In the whole-brain analysis, we also observed a significant encoding-encoding and interaction effects in anterior cingulate cortex, as well as recall-recall and interaction effects in dlPFC. These results suggest that both the "spoiled" manipulation and the "twist" may recruit top-down control and conflict monitoring processes during naturalistic viewing and recall."

    1. If I understand correctly, the main analyses of the fMRI data were limited to across-group comparisons of "critical scenes" that were maximally affected by the twist at the end of the movie. In other words, the analyses focused on the scenes whose interpretation hinged on the "doctor" versus "ghost" interpretation. I would be interested in seeing a comparison of "critical" scenes directly against scenes where the interpretation did not change with the twist. This "critical" versus "non-critical" contrast would be a strong confirmatory analysis that could further bolster the authors' claims, but on the other hand, it would be interesting to know whether the overall story interpretation led to any differences in neural patterns assigned to scenes that would not be expected to depend on differences in interpretation. (As a final note, such a comparison might provide additional analytical leverage for exploring the effect described in Figure 3B, which did not survive correction for multiple comparisons.)

    This is a helpful suggestion, and we’ve added an analysis addressing your comment. We found that the interaction index capturing the difference between the three groups was stronger for the critical scenes than for the non-critical scenes for almost all DMN ROIs.

    "The role of scene content In the prior analyses, we focused on “critical scenes”, selected based on ratings from four raters who quantified the influence of the twist on the interpretation of each scene (see Methods). An independent post-experiment analysis of the verbal recall behavior of the fMRI participants yielded “twist scores” that were also highest for these scenes; that is, the expected and perceived effect of twist information on recall behavior were found to match. In our next analysis, we asked whether the neural event representations reflect these differences in the twist-related content of the scenes. In other words, are the “critical scenes” with highly twist-dependent interpretations truly critical for our observed effects?

    To answer this question, we re-ran our main encoding-encoding and recall-recall pISC analysis in each DMN ROI (Figure 2-3). We calculated interaction indices (Figure 2C) first by including all scenes, and second by including only the 11 non-critical scenes. To better compare the effect of including different subsets of scenes to our original results, in Figure 4 we show the results in 15 ROIs that exhibited meaningful effects in our main analyses (Figure 2C). Figure 4A demonstrates that “critical scenes” yielded higher interaction indices compared to all scenes or non-critical scenes across all ROIs. The interaction score across all DMN ROIs was significantly higher in “critical scenes” than all scenes (t(23) = 7.19, p = 2.53 x 10-7) and non-critical scenes (t(23) = 7.3, p = 1.95 x 10-7). These results show that critical scenes are indeed responsible for the observed pISC differences across groups."

    1. I appreciate the code being made available and that the neuroimaging data will be made available soon. I would also appreciate it if the authors made the movie stimulus and behavioral data available. The movie stimulus itself is of interest because it was edited down, and it would be nice for readers to be able to see which scenes were included.

    Unfortunately due to copyright, we cannot share the movie stimulus outright. However, we will share the timing of the cuts used, as well as the time-stamped transcripts of verbal recall.

    To sum up, I think that this is a great experiment with a lot of strengths. The design is fairly clean (especially for a movie stimulus), the analyses are well reasoned, and the data are clear. The only weaknesses I would suggest addressing are with regards to how the DMN is being described and evaluated, and the communication of how this work informs the field on a theoretical level.

  3. Version published to 10.1101/2021.09.28.462068v3 on bioRxiv
  4. eLife

    Evaluation Summary:

    This is a well-designed study that will be of interest to cognitive neuroscientists studying event perception and memory, particularly those interested in naturalistic paradigms. The main contribution is in growing our understanding of how interpretational differences of events are reflected in differences in neural representations of those events. While the presented results are convincing, it remains somewhat unclear what processes drive the observed effects, and thus what the role of the implicated brain regions is in memory updating.

    (This preprint has been reviewed by eLife. We include the public reviews from the reviewers here; the authors also receive private feedback with suggested changes to the manuscript. Reviewer #3 agreed to share their name with the authors.)

  5. eLife

    Reviewer #1 (Public Review):

    This fMRI study investigated how memories are updated after reinterpreting past events. Participants watched a movie and subsequently recalled individual scenes from that movie. Importantly, the movie ends with a twist that changes the interpretation of earlier scenes in the movie. One group of participants watched the movie with the twist at the end, one group did not get to see the twist, and a third group was already informed about this twist before watching the movie. Analyses compared the similarity of activity patterns to (encoded or recalled) events across participants within regions of the default mode network (DMN). The design allowed for multiple relevant comparisons, confirming the prediction that activity patterns in DMN regions reflect the (re)interpretation of the movie (during movie viewing and/or during recall).

    The study is well-designed and executed. The inclusion of multiple analyses involving distinct comparisons strengthens the evidence for the role of the DMN in memory updating.

    The following points may be relevant to consider:

    1. The cross-participant pattern analysis method used here is not standard, with such analyses typically done within participants (or across participants, but after aligning representational spaces). Considering individual variability in functional organization, the method is likely only sensitive to coarse-scale patterns (e.g., anterior vs posterior parts of an ROI). This is not necessarily a weakness but is relevant when interpreting the results.

    2. Unlike previous work, analyses are not testing for scene-specific information. Rather, each scene is treated separately to establish between-group differences, and results are averaged across scenes. This raises the question of whether the patterns reflect scene-specific information or generic group differences. For example, knowing the twist may increase overall engagement, both when viewing the movie (spoiled group) and when recalling it (spoiled group + twist group). The DMN may be particularly sensitive to such differences in overall engagement.

    3. The study does not reveal what the DMN represents about the movie, such that its activity changes after knowing the twist. The Discussion briefly mentions that it may reflect the state of the observer, related to the belief about the identity of the doctor. This suggests a link to the theory of mind/mentalizing, but this is not made explicit. Alternatively, the DMN may be involved in the conflict (or switching) between the two interpretations.

    4. The design has many naturalistic aspects, but it is also different from real life in that the critical twist involves a ghost. Furthermore, all results are based on one movie with a specific plot twist. It is thus not clear whether similar results would be obtained with other and more naturalistic plot twists.

    5. Only 7 scenes (out of 18) were included in the analysis. It is not clear if/how the results depend on the selection of these 7 scenes.

  6. eLife

    Reviewer #2 (Public Review):

    In this manuscript titled "Here's the twist: How the brain updates the representations of naturalistic events as our understanding of the past changes", the authors reported a study that examined how new information (manipulated as a twist at the end of a movie) changes the neural representations in the default mode network (DMN) during the recall of prior knowledge. Three groups of participants were compared - one group experienced the twist at the end, one group never experienced the twist, and one group received a spoiler at the beginning. At retrieval, participants received snippets of 18 scenes of the movie as cues and were asked to freely describe the events of each scene and to provide the most accurate interpretation of the scene, given the information they gathered throughout watching.

    All three groups were highly accurate in the recall of content. The groups that experienced the twist at the end as well as at the beginning as a spoiler showed a higher twist score (the extent to which twist information was incorporated into the recall), while seemingly also keeping the interpretation without the twist ("Doctor representation") intact. Neurally, several regions in the DMN showed significant interaction effects in their neural similarity patterns (based on intersubject pattern correlation), indicating a change in interpretation between encoding and recall in the twist group uniquely, presumably reflecting memory updating.

    Several points that I think should be addressed to strengthen the manuscript:

    1. The results from encoding-retrieval similarity analysis (particularly the one depicted in Figure 3B) don't match the results from encoding/retrieval interaction (particularly those shown in Figure 2C). While they were certainly based on different comparisons, I would think that both analyses were set up to test for memory updating. Can the authors comment on this divergence in results?

    2. The recall task was self-paced. Can reaction time information be provided on how long participants needed to recall? Did this differ across groups? Presumably in the twist group and spoiled group participants might have needed a longer time to incorporate both the original and twist interpretation. How was the length difference across events taken into consideration in the beta estimates? Also, is there an order effect, such that one type of interpretation tended to be recalled first?

    3. The correlation analysis between neural pattern change and behavioral twist score is based on a small sample size and does not seem to be well suited to test the postulation of the authors, namely that some participants may hold both interpretations in their memory. Interestingly, the twist score of the spoiled group was similar to the twist group, indicating participants in this group might have held both interpretations as well. Could this observation be leveraged, for example by combining both groups (hence better powered with larger sample size), in order to relate individual differences in neural similarity patterns and behavioral tendency to hold both interpretations?

    4. Several regions within the DMN were significant across the analysis steps, specifically the angular gyrus, middle temporal cortex, and medial PFC. Can the authors provide more insights on how these widely distributed regions may act together to enable memory updating? The discussion on the main findings is largely at a rather superficial level about DMN, or focuses specifically on vmPFC, but neglects the distributed regions that presumably function interactively.

  7. eLife

    Reviewer #3 (Public Review):

    Zadbood and colleagues investigated the way key information used to update interpretations of events alter patterns of activity in the brain. This was cleverly done by the use of "The Sixth Sense," a film featuring a famous "twist ending," which fundamentally alters the way the events in the film are understood. Participants were assigned to three groups: (1) a Spoiled group, in which the twist was revealed at the outset, (2) a Twist group, who experienced the film as normal, and (3) a No-Twist group, in which the twist was removed. Participants were scanned while watching the movie and while performing cued recall of specific scenes. Verbal recall was scored based on recall success, and evidence for descriptive bias toward two ways of understanding the events (specifically, whether a particular character was or was not a ghost). Importantly, this allowed the authors to show that the Twist group updated their interpretation. The authors focused on regions of the Default Mode Network (DMN) based on prior studies showing responsiveness to naturalistic memory paradigms in these areas and analyzed the fMRI data using intersubject pattern similarity analysis. Regions of the DMN carried patterns indicative of story interpretation. That is, encoding similarity was greater between the Twist and No-Twist groups than in the Spoiled group, and retrieval similarity was greater between the Twist and Spoiled groups than in the No-Twist group. The Spoiled group also showed greater pattern similarity with the Twist group's recall than the No-Twist group's recall. The authors also report a weaker effect of greater pattern similarity between the Spoiled group's encoding and the Twist group's recall than between the Twist group's own encoding and recall. Together, the data all converge on the point that one's interpretation of an event is an important determinant of the way it is represented in the brain.

    This is a really nice experiment, with straightforward predictions and analyses that support the claims being made. The results build directly on a prior study by this research group showing how interpretational differences in a narrative drive distinct neural representations (Yeshurun et al., 2017), but extend an understanding of how these interpretational differences might work retrospectively. I do not have any serious concerns or problems with the manuscript, the data, or the analyses. However I have a few points to raise that, if addressed, would make for a stronger paper in my opinion.

    1. My most substantive comment is that I did not find the interpretive framework to be very clear with respect to the brain regions involved. The basic effects the authors report strongly support their claims, but the particular contributions to the field might be stronger if the interpretations could be made more strongly or more specifically. In other words: the DMN is involved in updating interpretations, but how should we now think about the role of the DMN and its constituent regions as a result of this study? There are a number of ideas briefly presented about what the DMN might be doing, but it just did not feel very coherent at times. I will break this down into a few more specific points:

    While many of us would agree that the DMN is likely to be involved in the phenomena at hand, I did not find that the paper communicated the logic for singularly focusing on this subset of regions very compellingly. The authors note a few studies whose main results are found in DMN regions, but I think that this could stand to be unpacked in a more theoretically interesting way in the Introduction.

    Relatedly, I found the summary/description of regional effects in the Discussion to be a bit unsatisfying. The various pattern similarity comparisons yielded results that were actually quite nonoverlapping among DMN regions, which was not really unpacked. To be clear, it is not a 'problem' that the regional effects varied from comparison to comparison, but I do think that a more theoretical exploration of what this could mean would strengthen the paper. To the authors' credit, they describe mPFC effects through the lens of schemas, but this stands in contrast to many other regions which do not receive much consideration.

    Finally, although there is evidence that regions of the DMN act in a coordinated way under some circumstances, there is also ample evidence for distinct regional contributions to cognitive processes, memory being just one of them (e.g., Cooper & Ritchey, 2020; Robin & Moscovitch, 2017; Ranganath & Ritchey, 2012). The authors themselves introduce the idea of temporal receptive windows in a cortical hierarchy, and while DMN regions do appear to show slower temporal drift than sensory areas, those studies show regional differences in pattern stability across time even within DMN regions. Simply put, it is worth considering whether it is ideal to treat the DMN as a singular unit.

    1. I think that some direct comparison to regions outside the DMN would speak to whether the DMN is truly unique in carrying the key representations being discussed here. I was reluctant to suggest this because I think that the authors are justified in expecting that DMN regions would show the effects in question. However, there really is no "null" comparison here wherein a set of regions not expected to show these effects (e.g., a somatosensory network, or the frontoparietal network) in fact do not show them. There are not really controls or key differences being hypothesized across different conditions or regions. Rather, we have a set of regions that may or may not show pattern similarity differences to varying degrees, which feels very exploratory. The inclusion of some principled control comparisons, etc. would bolster these findings. The authors do include a whole-brain analysis in Supplementary Figure 1, which indeed produced many DMN regions. However, notably, regions outside the DMN such as the primary visual cortex and mid-cingulate cortex appear to show significant effects (which, based on the color bar, might actually be stronger than effects seen in the DMN). Given the specificity of the language in the paper in terms of the DMN, I think that some direct regional or network-level comparison is needed.

    2. If I understand correctly, the main analyses of the fMRI data were limited to across-group comparisons of "critical scenes" that were maximally affected by the twist at the end of the movie. In other words, the analyses focused on the scenes whose interpretation hinged on the "doctor" versus "ghost" interpretation. I would be interested in seeing a comparison of "critical" scenes directly against scenes where the interpretation did not change with the twist. This "critical" versus "non-critical" contrast would be a strong confirmatory analysis that could further bolster the authors' claims, but on the other hand, it would be interesting to know whether the overall story interpretation led to any differences in neural patterns assigned to scenes that would not be expected to depend on differences in interpretation. (As a final note, such a comparison might provide additional analytical leverage for exploring the effect described in Figure 3B, which did not survive correction for multiple comparisons.)

    3. I appreciate the code being made available and that the neuroimaging data will be made available soon. I would also appreciate it if the authors made the movie stimulus and behavioral data available. The movie stimulus itself is of interest because it was edited down, and it would be nice for readers to be able to see which scenes were included.

    To sum up, I think that this is a great experiment with a lot of strengths. The design is fairly clean (especially for a movie stimulus), the analyses are well reasoned, and the data are clear. The only weaknesses I would suggest addressing are with regards to how the DMN is being described and evaluated, and the communication of how this work informs the field on a theoretical level.

  8. Version published to 10.1101/2021.09.28.462068v2 on bioRxiv
  9. Version published to 10.1101/2021.09.28.462068v1 on bioRxiv