Proceedings of the Annual Meeting of the Cognitive Science Society

Segmenting observations from an input stream is an impor-tant capability of human cognition. Evidence suggests that hu-mans refine this ability through experiences with the world.However, few models address the unsupervised developmentof event segmentation in artificial agents. This paper presentswork towards developing a computational model of how anintelligent agent can independently learn to recognize mean-ingful events in continuous observations. In this model, theagent’s segmentation mechanism starts from a simple stateand is refined. The agent’s interactions with the environ-ment are unsupervised and driven by its expectation failures.Reinforcement learning drives the mechanism that identifiesevent boundaries by reasoning over a gated-recurrent neuralnetwork’s expectation failures. The learning task is to reduceprediction error by identifying when one event transitions intoanother. Our experimental results support that reinforcementlearning can enable detecting event boundaries in continuousobservations based on a gated-recurrent neural network’s pre-diction error and that this is possible with a simple set of fea-tures.