Emergence of “Color Constancy” through Learning of Behaviors by Reinforcement Learning with a Neural Network

In this paper, “Optical Illusion” is considered as the result of unconscious process in our parallel and flexible brain, and the hypothesis is formed that it can be acquired through the learning to behave more appropriately. “Color constancy” is focused on and reinforcement learning is applied to a simple “colored-object guidance” task in which goal location depends on the object color with the condition that a translucent color filter covers a half of the field. It was observed that some hidden neurons came to represent the object color not depending on the filter color after learning. Furthermore, the neural network with new output neurons was trained to output the object color (RGB values) only under the condition of no filter, and then, when images covered by colored filter were provided as test inputs, the color represented by the network outputs was very close to the object color.