Proceedings of the Annual Meeting of the Cognitive Science Society

The event segmentation theory (EST) postulates that humanssystematically segment the continuous sensorimotor informa-tion flow into events and event boundaries. The basis for theobserved segmentation tendencies, however, remains largelyunknown. We introduce a computational model that groundsEST in the interaction abilities of a system. The model learnsevents and event boundaries based on actively gathered senso-rimotor signals. It segments the signals based on principles ofprobabilistic predictive coding and surprise. The implementedmodel essentially simulates, anticipates, and learns event pro-gressions and event transitions online while interacting withthe environment by means of dynamic, predictive Bayesianmodels. Besides the model’s event segmentation capabilities,we show that the learned encodings can be used for higher-order planning. Moreover, the encodings systematically con-ceptualize environmental interactions and they help to identifythe factors that are critical for ensuring interaction success.