AI tricked by illusionary motion of still images just like humans: study

Even artificial intelligence (AI) is tricked by optical illusions, a joint research group from the National Institute for Basic Biology, Ritsumeikan University and other organizations has discovered.

The study found that deep neural networks (DNNs), a kind of predictive AI based on structures in the human brain that employ deep learning techniques, mistakenly perceived illusionary motion like that of the stationary "rotating snake illusion" just like humans. The results were published in the international academic journal "Frontiers in Psychology" earlier this month.

The research group used software that could learn while modifying judgments based on visual information similar to what the human eye sends to the brain. The DNNs were trained using video of a rotating propeller to correctly estimate the direction, speed and other factors of motion. Then, when the AI was shown the still image of the "rotating snakes" optical illusion, it deemed that the circles within the design were actually moving, and calculated the direction and speed of the illusionary motion. When the color scheme of the design was changed, it was found that the DNNs perceived the rotation to be going in the opposite direction, as well as other reactions which matched that of humans viewing an optical illusion.

Deep learning AI has attracted attention in Japan, with the program "AlphaGo" defeating the world's top go players one after another. The technology can read a large amount of visual or other types of data, and learn from the examples to polish decision making and estimation proficiency skills on its own. The program is also used in clinical diagnostic imagining.

Team leader Eiji Watanabe, associate professor of animal psychology at the National Institute for Basic Biology, commented that the discovery of sensory illusions in DNNs have the potential to "contribute significantly to the development of brain research."