Click on the image to see a PDF version (for zooming in)

Fig. 9.3. Effect of environmental postnatal training patterns on orientation maps. Each simulation started with the same initial map, trained prenatally for 1000 iterations on noisy disks (ND) as shown in the top row of Figure 9.1. Postnatally, this map was trained for 9000 iterations under the same parameters but with retina-size segments of three different kinds of natural image inputs (the full images for these examples are shown in Figure 8.4d-f). In each case, maps with realistic features, RFs, lateral connections, and Fourier transforms developed. The final maps are less selective than those trained with artificial stimuli (Section 5.3), matching biological maps well. They also differ significantly on how the preferences are distributed. The network in the top row was trained on images of natural objects and primarily closerange natural scenes from Shouval et al. (1996, 1997). Like biological maps, this map is slightly biased toward horizontal and vertical orientations (as seen in the histogram), reflecting the edge statistics of the natural environment. The network in the second row was trained with stock photographs from the National Park Service (1995), consisting primarily of landscapes with abundant horizontal contours. The resulting map is dominated by neurons with horizontal orientation preferences (red), with a lesser peak for vertical orientations (cyan), which is visible in both the map plot and the histogram. The network in the bottom row was trained with upright human faces, by Achermann (1995). It has an opposite pattern of preferences, with a strong peak at vertical and a lesser peak at horizontal (bottom row). Thus, postnatal self-organization in HLISSOM depends on the statistics of the input images used, explaining why horizontal and vertical orientations are more prominent in animal maps, and how this distribution can be disturbed in abnormal visual environments. It also suggests that postnatal learning plays an important role in how visual function develops: It allows the animal to discover what the most important visual features are and allocate more resources for representing them.