Computational Maps in the Visual Cortex

ing the possible sources of ocular dominance, the LISSOM OD model above was based on inputs that differ in brightness in the two eyes (like the models of Bauer, Brockmann, and Geisel 1997; Riesenhuber, Bauer, Brockmann, and Geisel 1998). Some previous models have shown that OD maps might instead result from retinal disparity, i.e. differences in feature position in each eye (Burger and Lang 1999; Sirosh and Miikkulainen 1994a). Disparity itself is a separate feature dimension from ocular dominance. Neurons can be fully binocular yet prefer different positions in each eye, which is important for stereo vision. Even so, disparity does reduce the activity correlations between the eyes, and could thus contribute to ocular dominance as well. Disparity was modeled in LISSOM by randomly placing a spot at (x, y) in the left retina, and then placing a corresponding spot within a radius of ssR of (x, y) in the right retina, where R is the width of the retina. The scatter parameter ss ∈ [0..1] specifies the spatial correlations between spots in the two retinas. When ss = 0, the inputs are in perfectly matched positions, and when ss = 1 they are scattered independently (as in strabismus). By adjusting ss, it is possible to simulate different degrees of disparity between the images in the two eyes (see Appendix A.6 for the rest of the simulation parameters). In Figure 5.19, networks trained with brightness differences alone (ss = 0), with mild disparity (ss = 0.2), with moderate disparity (ss = 0.4), and with perfect strabismus (ss = 1.0) are compared. All these networks developed OD maps, but those with disparity are less realistic. With only mild disparity, the eye preferences 112 5 Development of Maps and Connections Left retina Right retina RFs LIs OD pref. OD H OD sel.