Predictive Inference Model of the Physical Environment that mimics Predictive Coding

Abstract

Research on real-world understanding, which represents humans' recognition and prediction functions in the physical environment as machine learning models, has been the focus of much attention in recent years. When humans look at the environment, it is believed that they recognize important timing and change points from a sequence of events. Based on the scenes they recognize, they may predict changes in events that are likely to occur in the future. In this study, we propose a model for predicting change points based on physical properties that mimic predictive encoding in the human brain. In the model, the physical properties of objects in the environment are represented by a graph structure. Therefore, the physical changes in the world can be predicted by predicting changes in the system. We verify that the expected timing correctly indicates significant change points such as collision or disappearance of objects.