Predictive Inference Model of the Physical Environment that emulates Predictive Coding

Eri Kuroda, Ichiro Kobayashi

Abstract: In recent years, the significance of artificial intelligence in comprehending the real-world has increased, by leveraging the inherent ability of humans to process intuitive physics on a computer. Prior investigations on real-world understanding have mainly relied on image inference to recognize the physical environment. In contrast, we propose an inference model that can predict the observed environment using both visual and physical features, emulating the predictive coding hypothesized to occur in the human brain, and detects change points in response to predictive events. Additionally, the model verifies the correctness of the timing of important physical events of objects, such as object collisions and disappearances. Furthermore, the results of the physical information prediction are also described as natural language sentences to confirm whether the model accurately recognizes the real-world and predicts the next behavior based on the physical information.