Face Localization - Hierarchical Neural Networks for Image Interpretation - page 205

Fig. 10.5. Face localization recall. The activities of the network's output feature arrays are
shown over time. Blobs indicating eye positions develop in a top-down fashion.
(a) (b) (c)
Fig. 10.6. Face localization recall. Shown are the contributions to the activity of the network's
output units after iteration 10 of Fig. 10.5 (bright shading represents inhibition, dark shading
indicates excitation): (a) via input and forward projections; (b) via lateral projections; (c) via
backward projections.
network's output is close to the desired one. It does not change significantly during
the next five iterations. Each iteration takes about 22ms on a Pentium 4 1.7GHz PC
without much optimization for speed.
The contributions to the network's output activities at iteration 10 are displayed
in Figure 10.6. It is evident that the main contribution to a blob's activity comes
via backward projections. They excite a larger area at the eye's position in all three
resolutions and inhibit its surround.
The effect of the lateral projections can be understood as center-center excitation
and center-surround inhibition. Note that the blobs do not exist independently, but
interact. This is most visible in the highest resolution, where both eye regions are
inhibited and only the center of the opposite eye is weakly excited.
The influence of forward projections and projections from the input feature ar-
rays is generally not that strong and less consistent. The input projections of Layer 0
seem to extract dark vertical lines that are surrounded by bright pixels from the im-
age. In Layer 2 and Layer 3, the forward and input projections excite the center of
the eye's blob and inhibit the blob of the opposite eye.
The network's dynamics cannot be understood by looking at the output feature
arrays alone. The majority of the computations are done by hidden features. Fig-
ure 10.7 shows their activities after ten iterations for the test example from Fig. 10.3.