Poster Session A, Saturday, March 23, 1:30 pm - 3:30 pm, Pacific Concourse
A latent-cause inference account of event segmentation under perceptual ambiguity
Yeon Soon Shin1, Yael Niv1,2, Sarah DuBrow1,3; 1Princeton Neuroscience Institute, Princeton University, 2Department of Psychology, Princeton University, 3Department of Psychology, University of Oregon
When experiencing a stream of events, we often segment them into meaningful clusters by inferring the underlying (latent) cause. Here, to characterize the process by which humans infer latent causes, we developed a novel perceptual categorization task and compared Bayesian models with and without temporal decay. We presented participants with “microbes” that varied along two dimensions (number and length of the microbes’ spikes) and had them infer the category of a microbe on each trial. Microbes belonged to one of four categories, where each category was generated from an underlying cause with a mean number and length of spikes. Sequential transitions mostly stayed within the same category, occasionally jumping to the next unexplored category (across-new) or going back to the previous category (across-revisit). Importantly, each across-category transition was paired with a within-category transition matched for perceptual distance. Participants (n=20) were more likely to infer a new category after an across-new transition than its distance-matched within-category pair, and more likely to infer an old than a new category after across-revisit transitions. A Bayesian clustering model with temporal decay best accounted for the categorization data, suggesting that humans adaptively use temporal contiguity when inferring the latent structure. In a subsequent memory test, when probed with each category’s prototypical (but novel) microbe, participants tended to choose the same category they had assigned to nearby microbes significantly above chance, indicating successful generalization. Together, these results highlight that temporal context sensitivity and generalization, processes linked to the hippocampus, may be important in inference-based event segmentation.
Topic Area: THINKING: Decision making