Neural Abstraction Pyramid Architecture - Hierarchical Neural Networks for Image Interpretation - page 90

Fig. 4.17. ZIP code binarization - network output. The network's stable foreground feature
responses to the examples from Figure 4.11(a) are shown.
(a)
(b)
Fig. 4.18. ZIP code binarization - results. (a) network input; (b) stable foreground feature
response.
line pixels to the foreground, although they are not much darker than their neighbor-
hood. Also, small clutter and serifs are removed from the foreground and assigned
to the background.
This kind of binarization improves the recognition of the ZIP code. In [24] I
showed that a similar network increases the acceptance rate of a ZIP code recogni-
tion system significantly without decreasing its reliability.
4.3.3 Activity-Driven Update
In Section 4.2.3 it was stated that the update of the cells in the Neural Abstraction
Pyramid occurs in a predetermined order: layer by layer in a bottom-up manner and
group by group within a layer. The following experiment investigates the effects of
relaxing the constraint of a predetermined update order.
This is motivated by the work of Thorpe et al. [226]. They found that the human
visual system is capable of rapid object categorization within 150ms. Thorpe and
Gautrais [227] proposed using a rank-order code, where a neuron emits at most one
spike to achieve such rapid feed-forward processing. Using this framework, Van-
Rullen et al. [232] showed that contour integration is possible even when neurons
fire asynchronously at most one spike. Integration of stimuli at less active neurons
is influenced by spikes emitted from more active neurons. This can facilitate or sup-
press the response to less salient stimuli.
The interpretation performance of the Neural Abstraction Pyramid depends on
the update order of its feature cells as Fig. 4.19 illustrates using a simple example. It
shows how the activity in a one-dimensional feature array develops in three different
update modes. The initial cell activities are set to the input stimulus, shown in the
front of the figure. It resembles a plateau that increases slightly from the edges (0.5)
towards the middle (0.56). The successive iterations show how the cell activities
develop over time under a dynamics described by: