Figure 1. A typical visual scene. The diameters of the receptive fields of the retina's input units are indicated by circles.
Figure 2. An artificial fovea provides inputs for a control network which is able to move the fovea around. A model network is trained to predict the next input from the current input and the current controller action.
Figure 3. By `substituting the model network for the environment' we obtain a recurrent combination of control network and model network. This new recurrent network is used for computing controller gradients by means of the `unfolding in time' algorithm.
Figure 4. Translations: Examples of fovea trajectories leading from various start positions to the target, which is the center of the crossing point in the letter `4'. No teacher told the fovea how to do that! Typically the system did not find the shortest path to the target. It developed a preference for edges.
Figure 5. One controller for various targets specified by an additional constant input: Examples of fovea trajectories leading from various start positions to different targets. The first target is near the left corner of the triangle. The second target is near the lower corner.
Figure 6. A triangle which may be arbitrarily rotated and translated in the pixel plane. The triangle is partly covered by some of the receptive fields of the moving fovea.
Figures 7 and 8. Examples of fovea trajectories leading from the outside of the object to the target. (For clarity, the fovea positions are indicated only for a fraction of all time steps.) No teacher told the system how to do that!
Figure 9. The fovea pushing backwards to the target on a noisy pixel plane.