Learning Iterative Image Reconstruction - Hierarchical Neural Networks for Image Interpretation - page 192

Fig. 9.23. MNIST reconstruction from sequence of degraded images. The activities of all
feature arrays are shown over time.
(a)
(b)
(c)
(d)
(e)
Fig. 9.24. Reconstruction from sequence of degraded images. For the example of Fig. 9.23
are shown over time: (a) hidden feature array in Layer 0 (representing line-candidates); (e)
output feature array (the reconstructed digit); contributions to the output activity (b) via input
projections (noisy); (c) via lateral projections (center-surround interaction); (d) via backward
projections (inhibiting the background stronger than the lines).
in the bottom layer, need more time steps than in the previous experiments to reach
a high confidence level. Towards the end of the sequence, the background clutter
and the occluding square have been removed by the network from the output fea-
ture array, and the lines have been completed. The network's activities converge to
a distributed representation of the digit which facilitates the reconstruction of the
original.
In Figure 9.24, the activities of the single hidden feature array in Layer 0 and
the output feature array, together with its contributions are shown for the same digit
over time. The hidden feature is more active than the output feature, representing
line candidates. Background structures are highly visible in the first few iterations.
They are reduced towards the end of the sequence.