Neural network model of color vision

摘要

A three-layer neural network model for color vision is constructed on the basis of physiological evidence such as the spectral response properties of cone and color-coded cells in V4 and trained by using a back-propagation learning algorithm. Six types of chromatic response properties of hidden units were constructed by a network learning to transform the broadband color space to the narrowband color space. The results show that the response properties play an essential role in color vision and neural network capability, providing an effective method for elucidating higher neural mechanisms such as representation of information or information coding in neural circuits. It was found that trained learned hidden units have characteristics similar to those of the color cells found in macaque lateral geniculate nucleus.