Predictive networks generate motion-induced color illusions.

Subjective color (SC) refers to chromatic perception induced by achromatic stimuli, as classically demonstrated using Benham's top. Despite nearly two centuries of investigation, the underlying neural mechanisms remain unclear. We developed artificial neural network models trained on natural videos using predictive learning to investigate SC generation. These models successfully reproduced artificial subjective color (ASC) from achromatic testing stimuli that contained no explicit chromatic information. Critically, ASC characteristics were systematically influenced by the colors of moving objects in the training videos. Using progressively simplified stimuli, ranging from natural scenes to 3D computer graphics and 2D animations, we demonstrated that the colors of moving objects primarily determine ASC properties. We propose that predictive learning establishes color associations with motion patterns, causing the arcs on Benham's top to be perceived as moving objects with corresponding colors. These findings suggest that cortical predictive mechanisms may complement retinal processes in generating subjective color phenomena.