How attention simplifies mental representations for planning
- Department of Experimental Psychology and Institute for Cognitive Neuroscience, University College London, United Kingdom
- Institute of Philosophy, School of Advanced Study, University of London, United Kingdom
- Max Planck UCL Centre for Computational Psychiatry and Ageing Research, University College London, United Kingdom
- Canadian Institute for Advanced Research (CIFAR), Brain, Mind and Consciousness Program, Canada
Figures
Figure 1 with 12 supplements
Spatial attention shapes task representations.
(a) Schematic of the maze navigation task. Participants fixated at the start of each trial, after which a maze was presented, which they were asked to navigate. Maze stimuli either remained on the screen during navigation (left panel; concurrent planning experiments) or were removed before navigation (right panel; upfront planning experiments). Once participants finished navigating the maze, they were asked to report their awareness of every obstacle presented on a given trial in a random order. (b) Left panel: schematic of the analysis pipeline. An example maze is shown where seven obstacles (plotted in orange) are presented on every trial according to pre-defined mazes. Participants report their awareness of every obstacle at the end of each trial (middle maze). The value-guided construal (VGC) model predicts which obstacles in a maze will likely be included in participants’ task representation (right maze). We use participants’ awareness reports to test the influence of neighbouring obstacles on the probe obstacle (presented in green). We compute the influence of neighbouring obstacles (in grey) on participants’ awareness of the probed obstacle (in green). Right panel: Results of the ranked regression model for dataset Ho 1. We observed that obstacles closest to the probed item (ranks 1 and 2) positively impact awareness reports. In contrast, obstacles furthest from the probed item negatively impact awareness reports (ranks 5 and 6). (c) Left panel: The effect of neighbouring obstacles on task representations varied across participants (each represented by a point). Right panel: Inter-individual differences in the attentional effects correlate with the sparsity of participants’ representations. Participants who showed the greatest influence of neighbouring obstacles (more negative slopes) showed the simplest representations (greatest variance in awareness reports).
Figure 1—figure supplement 1
Model predictions and reported awareness on mazes 0–5 of reanalysed data.
Figure 1—figure supplement 2
Model predictions and reported awareness on mazes 6–11 of reanalysed data.
Figure 1—figure supplement 3
Model predictions and reported awareness on non-lateralised mazes 0–5 of the new experiment.
Figure 1—figure supplement 4
Model predictions and reported awareness on non-lateralised mazes 6–11 of the new experiment.
Figure 1—figure supplement 5
Model predictions and reported awareness on lateralised mazes 0–5 of the new experiment.
Figure 1—figure supplement 6
Model predictions and reported awareness on lateralised mazes 6–11 of the new experiment.
Figure 1—figure supplement 7
Spatial proximity predicts awareness for both task-relevant and -irrelevant obstacles.
Standardised beta coefficients of the ranked regression model for task-relevant and -irrelevant obstacles separately across datasets Ho 1 and 2. We observed that obstacles closest to the probed item (ranks 1 and 2) positively impact awareness reports, regardless of task-relevance. In contrast, obstacles furthest from the probed item negatively impact awareness reports (ranks 5 and 6).
Figure 1—figure supplement 8
Neighbouring obstacles predict inclusion/exclusion from task representation.
Right panel: Results of the ranked regression model for the upfront planning experiment (i.e. dataset Ho 2). Obstacles closest to the probed item (ranks 1 and 2) positively impact awareness reports, whereas obstacles furthest from the probed item negatively impact awareness reports (ranks 4, 5, and 6). Left panel: Inter-individual differences in spatial attention effects in dataset Ho 2. Each participant is represented by a point and line.
Figure 1—figure supplement 9
Spatial proximity predicts awareness for dataset dSC 1.
Standardised beta coefficients of ranked regression model for experiment 3. We observed that obstacles closest to the probed item (ranks 1 and 2) positively impact awareness reports, regardless of the type of maze (i.e. lateralised vs non-lateralised; right panel). In contrast, obstacles furthest from the probed item negatively impact awareness reports (rank 5). We note that while the 2nd closest obstacle positively predicted the awareness reports, this effect was much weaker than the closest obstacle (i.e. rank 1).
Figure 1—figure supplement 10
Inter-individual differences in simplified representations.
(a) Analysis pipeline to explore the relationship between spatial attention and task-representations. First, we fit a linear model to beta coefficients obtained in Figure 1b left for each participant. This slope corresponds to a participant’s attention effect, with more negative slopes indicating larger attention effects. We then computed the sparsity of participant’s representations by computing the mean variance of their awareness reports. (b) Examples of participants’ task representations. The top row depicts the average awareness of each obstacle in three example mazes. The second row depicts the representations of participants with the most negative attention slopes. The bottom row plots the awareness effects of participants with the shallowest slopes (i.e. weakest spatial attention effects). (c) Participants with stronger spatial attention effects had more sparse task representations. We replicated this effect in two independent samples across datasets Ho 2 and dSC 1.
Figure 1—figure supplement 11
Inter-individual in attentional spotlight.
Right panel: We investigated inter-individual differences in participants’ attentional spotlight width. To do so, we defined the width of their attentional spotlight as the rank at which their spatial proximity beta coefficient – computed for lateralised mazes – went from strongly positive to below 0.05. We used this transition point to define the rank distance that represented the shift between a positive to a negative influence of neighbouring obstacles on a probe obstacle’s awareness (see Figure 2). We computed the median distance between obstacles at this rank for every participant and used this value as the width of their attentional spotlight (see left panel).
Figure 1—figure supplement 12
Robustness of the attentional spotlight model to spatial autocorrelation.
We assessed the robustness of the spotlight model to spatial autocorrelation by performing two sets of different spatial permutations. (a) In the first set of permutations, we permuted the predictions of the spotlight model across all mazes and used these null predictions in a hierarchical linear regression model. We repeated this procedure 1000 times to produce a null distribution of beta coefficients, depicted in the right panel. The observed spotlight model effect (dotted line) was significantly better than the spatial null permutations. (b) We similarly permuted the spotlight model’s predictions within each maze, assigning each obstacle a random prediction. We used these null predictions in a hierarchical regression model to predict participants’ awareness reports. We repeated this procedure 1000 times to generate a null distribution, depicted in the right panel. The spotlight model predicted participants’ awareness reports beyond the spatial autocorrelation of the data.
Figure 2 with 2 supplements
Lateralisation of task-relevant information affects task representations.
(a) For each maze, we computed a vertical meridian lateralisation index. This index reflects whether task-relevant information is lateralised to a hemifield. In the example plotted, there is more task-relevant information presented on the left than on the right of the maze, therefore this would correspond to a moderate level of vertical meridian (i.e. left vs right) lateralisation. We similarly computed an attention index for the horizontal meridian (i.e. above vs below). (b) Density plots of the reported awareness of obstacles on the basis of whether the value-guided construal (VGC) model predicted them to be task-relevant (≥0.5; in orange) or task-irrelevant (<0.5; in grey). Note sVGC model predictions for each obstacle were binarised for visualisation purposes only. Participants were more likely to be aware of obstacles predicted as task-relevant. We split maze stimuli based into terciles based on the degree to which task-relevant information was presented preferentially to one hemifield (x-axis). The leftmost plots are mazes where task-relevant information is presented on both hemifields. In contrast, the rightmost plot depicts mazes with the largest lateralisation. We observed that the awareness reports of participants become increasingly aligned to the VGC model’s predictions as lateralisation increases. (c) Scatter plot of the effect of maze lateralisation on the relationship between the value-guided model and participants’ awareness of obstacles. We observed a significant vertical meridian lateralisation effect whereby participants’ awareness reports were more strongly aligned with the VGC model’s predictions when task-relevant information was presented unilaterally in all datasets. Each point represents an obstacle in a maze, and each line represents the model fit for that specific maze stimulus.
Figure 2—figure supplement 1
Eye position during planning.
We verified that our behavioural effects were not driven by eye movements during the planning phase in dataset dSC 1. (a) The fluctuations of eye position, measured as the SD of the eye-position time series, did not statistically differ between non-lateralised and lateralised maze stimuli. (b) Time series of the average position of participants’ gaze during planning along the horizontal and vertical axes, shown in left and right panels, respectively. Participants, on average, moved their eyes more toward the left when maze stimuli were lateralised to the left (yellow line) and toward the right when the maze stimuli were lateralised to the right (orange line). These eye movements, however, remained within the bounds of the central square (dotted grey lines). Eye gaze did not differ between maze stimuli along the y-axis. (c) Heat maps depicting where participants looked during the planning phase for non-lateralised, right-lateralised, and left-lateralised maze stimuli.
Figure 2—figure supplement 2
Attention lateralisation effects are robust to eye-gaze.
Density plots of the reported awareness of obstacles separated task-relevant (≥0.5; in orange) or task-irrelevant (<0.5; in grey) obstacles as predicted by the value-guided construal (VGC) model for trials with minimal eye movements. Participants were more likely to be aware of task-relevant obstacles and unaware of irrelevant obstacles. This effect was moderated by the degree to which task-relevant information was presented preferentially to one hemifield (x-axis). From left to right, we plot the three terciles of maze lateralisation. Participants’ awareness reports become increasingly aligned with the VGC model’s predictions – that is the overlap between the two density plots decreases with lateralisation.
Figure 3
Inter-individual variation in lateralisation of awareness.
(a) For each maze and participant, we computed an awareness lateralisation index (ALI). This index reflects the degree to which participants tended to pay attention to obstacles in a single hemifield. In the example plotted, the participant preferentially paid attention to the obstacles presented to the left hemifield regardless of whether they were task-relevant or task-irrelevant. Note that this lateralisation index is based on participants’ self-reports, unlike the lateralisation index presented in Figure 2 which is based on the sVGC model predictions. (b) Histogram of the ALI of participants across the Ho 1 and 2 datasets. In these experiments, some participants showed substantial lateralisation of awareness (ALI >0.5), despite the maze stimuli for these experiments being – on average – non-lateralised in their sVGC model predictions. (c) Histogram of the ALI of participants for maze stimuli with non-lateralised VGC model predictions. We plot ALI values separately for the original non-lateralised mazes, and the left-right reversed (flipped) mazes separately. Participants on average did not lateralise their awareness. We note, however, that on some trials participants’ awareness reports were strongly lateralised, which contrasts with the sVGC model predictions. (d) Scatter plot of participants’ tendency to lateralise their attention to either hemifield (i.e. absolute value of ALI). We plot this for mazes with left and right lateralised model predictions (left panel) and for mazes with non-lateralised and lateralised model predictions (right panel). The large linear relationships indicate that participants’ tendency to lateralise their awareness is a stable inter-individual difference.
Figure 4 with 1 supplement
A value-guided construal (VGC) model augmented with an attentional spotlight model predicts participants’ task representations.
(a) Schematic of the attentional spotlight model. Inspired by the spotlight of attention analogy, we recompute an obstacle’s probability of being included in a task representation as the weighted average of its neighbours. We first search for all neighbours of obstaclei that are w squares away. We then compute P(Obstaclei) as the weighted average of obstaclei and its neighbours. This generates more graded model predictions (far right panel). (b) Left panel: Each row represents a different example maze stimulus. The left column depicts the original VGC model prediction P(Obstaclei) for every obstacle in the example maze. The middle column shows the attentional-spotlight model prediction for every obstacle. Obstacles that were considered task-relevant (deep orange) in the original model become less important when surrounded by task-irrelevant information (grey obstacles). The right column shows the participants’ average awareness of each obstacle in the example mazes. Right panel: Scatter plot of the linear relationship between participants’ awareness reports of obstacles and model predictions (original-VGC in green and the spotlight-VGC model in orange) for dataset Ho 1. The latter fits participants’ reports better than the original VGC model. (c) Scatter plot of the linear relationship between participants’ awareness reports of obstacles and model predictions (original VGC in green and the spotlight-VGC model in orange) for dataset dSC 1 separately for non-lateralised (left panel) and lateralised mazes (right panel). Although both models fit participants’ awareness reports better for lateralised mazes, the advantage of the spotlight model over the original model (better model fit / lower BIC) was observed only in non-lateralised mazes.
Figure 4—figure supplement 1
Distribution of distances between obstacles for each rank in regression.
Boxplots depicting the distance between obstacles. In datasets Ho 1 and 2, obstacles ranked 1st (closest) and 2nd in proximity were on average 2.14 (median = 2.0, sd = 0.91) and 3.60 squares away (median = 3.0, sd = 1.30), respectively. These obstacles positively predict participants’ awareness of items. In dataset dSC 1, obstacles ranked closest were on average 2.76 (median = 3.0, sd = 0.83) squares away. Based on these statistics, we expected an attentional spotlight of width 3 would provide the best fit to the data.
Additional files
-
Supplementary file 1
Control analyses and effects of neighbouring obstacles.
- https://cdn.elifesciences.org/articles/108034/elife-108034-supp1-v1.xlsx
-
Supplementary file 2
All linear regression models.
- https://cdn.elifesciences.org/articles/108034/elife-108034-supp2-v1.xlsx
-
MDAR checklist
- https://cdn.elifesciences.org/articles/108034/elife-108034-mdarchecklist1-v1.pdf
Download links
A two-part list of links to download the article, or parts of the article, in various formats.
Downloads (link to download the article as PDF)
Open citations (links to open the citations from this article in various online reference manager services)
Cite this article (links to download the citations from this article in formats compatible with various reference manager tools)
How attention simplifies mental representations for planning
eLife 14:RP108034.
https://doi.org/10.7554/eLife.108034.3
