4.1 Training inputs

  To develop an orientation map in the RF-LISSOM model, one of the simplest oriented training patterns was used: a 2-dimensional elongated Gaussian. For these inputs, the activity at each retinal ganglion cell is calculated according to the equation

 $$\begin{eqnarray} \html{eqn8} \xi_{r_1,r_2} & = exp(& {-\frac{((r_1-x_i)cos(\al... ... & {-\frac{((r_1-x_i)sin(\alpha)+(r_2-y_i)cos(\alpha))^2}{b^2} }) \end{eqnarray}$$

where a² and b² specify the length along the major and minor axes of the Gaussian, (x_i,y_i) specifies its center, and $\alpha$its orientation. The x and y coordinates of the centers are each chosen randomly within the retinal area, and the orientation is chosen randomly from the uniform distribution in the range 0° $\leq$$\alpha$ $<$ 180°. Figure 4.1 shows several sample training inputs.

 

Figure 4.1: Sample training inputs plotted on the retina.

These patterns approximate natural visual stimuli after the edge detection and enhancement mechanisms in the retina. Such edge-enhanced images have a predominance of elongated features. Similar features may also be found in the intrinsic retinal activity waves that occur in late prenatal development in mammals, and they are believed to drive the initial organization of the visual cortex (as discussed in section 3.7.4; Catsicas and Mobbs 1995; Meister et al. 1991; Wong et al. 1993). The RF-LISSOM network models the self-organization of the visual cortex based on these natural sources of elongated features.