Emergence of color constancy illusion through reinforcement learning with a neural network

Our parallel and flexible brain that must be the origin of our flexibility processes visual signals without being noticed, and due to the unawareness, the contradiction between our perception after the process and original visual property is exposed as “Optical Illusion”. The authors form the hypothesis that optical illusion can be acquired through or supported by the learning so as that we behave more appropriately in everyday life. In this paper, “color constancy” is focused on and the authors try to explain its emergence through the learning of a simple “colored-object guidance” task by reinforcement learning with a neural network whose inputs are raw image signals. In the task, it is required to move an object whose color is chosen randomly to the proper location that differs depending on the object color. Half of the field is covered by a translucent filter whose color and angle are chosen randomly at each episode. It was observed that the hidden neurons came to represent the object color mainly not depending on the filter color after reinforcement learning. In the subsequent supervised learning and test, the neural network with new output neurons was trained to output the object color only under the condition of no filter, but, when images covered by colored filter were the input as test patterns after learning, the network outputs were very close to the original object color.