In lab, AI generates faint colors from a black-and-white spinning top

Spin a black-and-white Benham’s top, and faint colors can seem to appear where none were drawn. Kyohei Ueda, a graduate student at the Laboratory of Neurophysiology at the National Institute for Basic Biology, and his colleagues have now made an artificial neural network produce a similar effect in its predictions.

The model was trained through predictive learning: it watched natural videos and learned to predict what kind of image would come next. When the researchers showed it the black-and-white pattern, faint colors emerged in the images it generated, despite the absence of color in the input.

The effect was not fixed. The colors produced by the model changed depending on the colors of moving objects in the videos used for training. Tests with natural videos, 3D computer graphics and simple two-dimensional videos of moving red, green and blue squares pointed to learned associations between motion and color as one factor behind the illusion.

The result adds a new possibility to a puzzle studied for approximately 200 years. Subjective color may not be created solely within the retina; it may also reflect the brain predicting future visual input from past experience. The study, published in Scientific Reports, is a model-based laboratory investigation, so it does not yet prove that human vision uses exactly the same mechanism.

So what changes concretely? Researchers gain an artificial system in which motion-color links can be adjusted and observed while studying a classic visual illusion. For readers, the result shifts the explanation from a purely retinal effect toward a process that may include learned prediction—but the evidence remains limited to the model and the experiments described.