Learning Iterative Image Reconstruction - Hierarchical Neural Networks for Image Interpretation - page 184

Fig. 9.13. Filling-in of occluded parts. The activities of all feature arrays are shown over
time. All features contribute to the computation. The activities change most in the first few
iterations. Towards the end of the sequence, the network approaches an attractor that includes
the reconstructed digit in the output feature array.
Input 3 6 12 Target Input 3 6 12 Target
Fig. 9.14. Filling-in of occlusions. The activities of the network's outputs are shown over
time for the first ten test images. The recurrent network is able to remove the square, and it
produces a reasonable guess of the digit's appearance.