Abstract

Ahumada and Mulligan¹ proposed a network model for constructing a red–green opponent system from LGN outputs without specific long versus middle wavelength cone labeling. They constructed model LGN cells having long and middle cone inputs and fed the LGN outputs into units, presumed cortical, which were trained by a network learning process to compute the principal component of the LGN units in its receptive field. These units turn out approximately doubly opponent and have less luminance sensitivity than their LGN input cells. Their outputs were then calibrated by translation invariance. Derrington, Krauskopf, and Lennie² and Young³ have published measurements of relative long, middle, and short wavelength cone weightings for actual LGN cells. These weights are sufficient to compute LGN outputs for training the model opponent cells. Model cortical cells were constructed by taking a random sample of seven LGN cells from a data set and computing their principal component weights. The resulting cells are highly red-green opponent (less luminance sensitivity and smaller short wavelength cone weightings). The strong opponency of the LGN cells and the infrequency of large short wavelength weights allows the construction of red-green opponent cells without knowledge of their input sources.

© 1990 Optical Society of America