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

Fig. 9.1. Effect of internally generated prenatal training patterns on orientation maps. Three different networks were trained for 1000 iterations to match newborn orientation maps as well as possible. The networks and training parameters were otherwise identical except different training inputs were used. As in Figure 5.13, the columns show a sample retinal activation, the LGN response to that activation, self-organized receptive fields for sample neurons, lateral inhibitory weights of these same neurons, the organization of the orientation map with selectivity superimposed in gray scale, and the histogram and the Fourier transform of the OR preferences. Overall, the features seen in the corresponding fully organized maps of Figure 5.13 have already started to emerge in each of these maps, although they are less distinct at this stage. They contain linear zones, pairs of pinwheels, saddle points, and fractures, and their retinotopic organization and gradient (not shown) are roughly similar to adult maps. The ring-like shape of the Fourier transform is also starting to emerge with disk and noisy disk inputs. The map obtained with noisy disks is the best match with animal maps (Figure 9.2). Note that nearly all of the resulting receptive fields have two lobes (i.e. they are edge-selective) rather than three (lineselective), predicting that a similar distribution would also be found in newborns. With noiseless patterns (middle row), the RFs are very smooth, and the neurons become highly selective for orientation, unlike neurons seen in newborn maps. On the other hand, with uncorrelated random noise (bottom row), the neurons become significantly less selective and the RFs do not have regular shapes like they do in animals. The "Noisy disks" map therefore constitutes the most realistic model of prenatal self-organization, and will be used as a starting point for postnatal training.