Event Segmentation Theory (EST) postulates that humans systematically segment the continuous sensorimotor information stream into events and event boundaries. We present a computational model that is able to learn predictive encodings of events and event transitions online from the sensorimotor information flow perceived by an embodied agent while exploring its own behavioral capabilities and interacting with its environment. Our system currently explores various behavioral modes by means of a self-organizing artificial neural network-based controller, which changes its internal weights based on a biological plausible synaptic plasticity rule, generating motor commands that lead to the emergent generation of various rhythmic behavioral dynamics. While exploring its behavioral capabilities our system uses a set of forward models, i.e. event models, that anticipate the ongoing sensorimotor changes. Statistical analysis of the unfolding predictions is used to detect significant prediction errors, which we use as an indicator for an event boundary and a transition in forward models. In this way our system learns compositional encodings of events and event boundaries of sensorimotor time series completely unsupervised. Furthermore, our system is able to actively use learned event models for reproducing behavioral patterns and planning goal-directed behavior to reach a desired goal state.
Last modified: Wednesday, 28-Feb-2018 13:09:30 NZDT
This page is maintained by the seminar list administrator.