Color opponency as the internal representation acquired by a three-layered neural network model

The color representation in the visual system is discussed through the analysis of a three-layered neural network model incorporating physiological evidence of color representation at the sensor level and the perceptual level. The model is trained to perform a mapping between these color representations by a backpropagation algorithm. The acquired characteristics of the hidden units are analyzed. The hidden units learn characteristics similar to those of the color opponent cells found in fish retina and macaque lateral genticulate nucleus (LGN). It is concluded that the R-G and Y-B color opponent representations play an essential role in color information processing by investigating the efficiency of color representation in the hidden layer and the capability of color discrimination task of the model. >