Analysis of Factors Causing Illusions in Next Frame Prediction Model Based on Predictive Coding Theory

Abstract

Illusions play a crucial role in how we perceive reality, and understanding how they occur aids in studying human visual processing. Recently, a model named PredNet, inspired by the brain's predictive coding, was trained using videos reflecting a human-like perspective of the world. Interestingly, without specific training on illusions, PredNet has shown to generate predictions resembling human motion illusions. In our study, we trained PredNet with datasets featuring various observer movements to see if it would produce illusion-like predictions, aiming to understand the learning aspects in predicting illusory motion. We found that PredNet, trained with different datasets, made notably distinct predictions. This indicates that the movement of the observer significantly influences the emergence of illusions and the perception of the world.