Fig. 2: Viewing behavior is event-specific and consistent across participants.
From: Neural and behavioral reinstatement jointly reflect retrieval of narrative events
A Group-level gaze example. Horizontal gaze position averaged across participants (green line) with standard deviation (shaded area) for Dataset 1 (left) and Dataset 2 (right). Gaze behavior was highly consistent across participants and datasets. B Saccade parameters: Scatter plots show saccade rate, amplitude, and duration for each narrative event averaged across participants (green dots) for both datasets. Regression lines added. Saccade parameters vary across events, and this variability is consistent between datasets. C Event-by-event gaze-map similarity for Dataset 1 (left) and Dataset 2 (middle), as well as for model-derived saliency54 of the movie stimulus (right). Matrices show the ranks of Pearson correlations between gaze maps obtained for each event, where a gaze map is defined as a 2D histogram of gaze positions aggregated over the duration of an event. Gaze-map similarity thus reflects the spatial similarity in fixation patterns between events. Ranking was performed after computing the correlations to normalize matrix range (blue to yellow colors show low to high ranks). Saccades and gaze positions were event specific and robust across datasets. D Event-specific gaze patterns are highly reliable and explained by frame-wise saliency, not sentence embeddings. Scatter plots show the relationship between the lower diagonals of the event-by-event similarity matrices in (C) and Fig. 1D: Eye tracking (ET) Dataset 1 vs. 2 (left, r = 0.73, p = 6.9 × 10-185), movie saliency vs. averaged eye tracking (middle, r = 0.64, p = 6.8 × 10-130), averaged sentence embeddings (sBERT) vs. averaged eye tracking (right, r = 0.1, p = 1.3 × 10-3). Regression line added.
