The Japanese Society for Artificial Intelligence

Human vision does not faithfully reproduce physical parameters of visual objects. One of the reasons for this is that only a limited amount of visual information is available to the brain, and the brain estimates the visual world from this ambiguous information. However, the reproduced visual world, although "inaccurate," has very useful information for the people who live there. The predictive coding is one of the theories to explain such a visual processing of the brain. At the core of the theory is the “prediction” concept, which states that the brain is always making predictions, and that prediction errors drive learning. Recently, our group has been trying to reproduce visual illusions using deep learning networks that incorporates the predictive coding. Visual illusion is a representative example of a significant discrepancy between perceptual and physical parameters, which can be used as a criterion to verify the theory. In this talk, I would like to introduce a new style of perception research by showing that many illusions can be reproduced by deep learning models, and that it is possible to create illusions by combining deep learning models and genetic algorithms.