Fig. 1: One-shot learning of event boundaries and word predictions.

a Agreement between raters on event boundaries. Blue and red lines depict the ratio of raters that marked an event boundary within every second in the story on run 1 and run 2. Lines in fuchsia indicate boundaries based on run 2, as per peak detection. b Similarity of each subject’s response vector to the agreement between all others. Purple dots indicate individuals: subjects above the diagonal increase in similarity to others on the second run. The average increase (red cross) indicates consensus learning. Source data are provided as a Source Data file. c Zoomed-in time interval of agreement between raters (compare panel a). Agreement on the second run (red line) slightly precedes agreement on the first run (blue line). d Earlier boundary detection on run 2 is marked by a negative lag in the cross-correlogram (purple), peaking at −182 ms (turquoise vertical line). e Performance difference between groups that have listened to the story and naive participants in prediction of upcoming words in the story (the naive groups were  predicting only based on general knowledge of language, lacking episodic information about the narrative). Predictive recall of upcoming words manifests itself in a positive difference between the groups, i.e., prediction probability of the words increases after a single exposure (turquoise prediction experiment, fuchsia replication; the prediction experiment was replicated once in a new sample), demonstrating one-shot learning of story content. Notably, there is substantial variance across words, suggesting that some words are learned better than others.