Information Technology Reference
In-Depth Information
The contributions from the input to the output feature array are weakly exci-
tatory with a low-pass characteristic. Input noise, the background level, and the
occluding square are highly visible here.
The contributions from lateral projections are strongly excitatory in the center
and weakly inhibitory in the surroundings. Hence, lines excite themselves and their
immediate neighborhood, and inhibit their surround.
The contributions via backward projections seem again to inhibit the output fea-
ture according to the estimated background level. Interesting is the strong inhibition
of the units near the image border at iteration three. This is the first step where in-
formation from Layer 3, which has a global view of the image, reaches the output.
Towards the end of the sequence, the inhibition is weaker at the units belonging to
lines. This shows that lines are represented at higher layers.
Figures 9.25, 9.27, and 9.29 display the reconstruction process of the first ten
test digits for low noise, medium noise, and high noise, respectively.
One can see that in all three cases the network is able to produce good recon-
structions of the originals, which are also shown in the figures. The less ambiguous
image parts are reconstructed faster than the parts that are occluded. The higher the
noise is or the stronger the background level deviates from the mean, the more iter-
(a) 1 2 4 7 11 16 (b) 1 2 4 7 11 16 (c)
Fig. 9.25. Reconstruction from a sequence of degraded MNIST images (low noise): (a) input
over time; (b) output over time; (c) target.
Search WWH ::




Custom Search