Page 107 - Biomimetics : Biologically Inspired Technologies

Bar-Cohen : Biomimetics: Biologically Inspired Technologies DK3163_c003 Final Proof page 93 21.9.2005 11:40pm

Mechanization of Cognition 93

Figure 3.11 Object segmentation example. A portion of a wide-angle camera frame is shown. The gaze
controller has fixated upon the upper left back corner of a rectangular solid (shown here in color for clarity). The
eyeball image is shown surrounding the fixation point with the 81 (overlapping — they actually overlap a bit more
than is shown here) fields of view of the 81 primary visual lexicons shown. As explained in the text, the primary
visual layer knowledge bases can be used to eliminate the lexicon responses to all visual objects except the one
upon which the fixation point lies.

the eyeball image have all of their symbols shut down and thereby become null (this follows from
the fact, discussed in Section 3.1, that the only symbols of a lexicon with a frozen expectation which
can receive input excitation from a knowledge link are those which belong to the expectation). In
general, the only way that the expectation of an outlying lexicon can have any symbols retained is if
one or more of its expectation symbols codes a local appearance that has been meaningfully seen
before in conjunction with one or more of those expectations of the lexicons proximal to the fixation
point.
A more elaborate version of this process can also be used, in which a ‘‘wave’’ of confabulations
moves outward from the middle of the primary lexicon array to the periphery; with only knowledge
bases spanning one or two inter-lexicon distances being enabled as the wave progresses. This
improves performance because closer-distance related appearances are more likely to have
appeared enough during training to be considered meaningful and be retained.
The astounding thing about this process (which is very fast because all of the distal lexicon
confabulations happen in parallel) is that it effectively SEGMENTS THE OBJECT UPON
WHICH THE FIXATION POINT LIES from all the other image content of the eyeball
image. In other words, ideally, after this segmentation procedure, which is virtually instanta-
neous, only symbols describing local appearance of the attended object (the one selected by the
gaze controller having the fixation point sitting on it) remain in the expectations of the primary
visual layer lexicons. In Figure 3.11, these nonnull lexicons (representing the rectangular
solid shown) are illustrated as diagonally-hatched in magenta. In other words, the only visual
appearance data left on the primary visual layer is that describing the attended object, which
has thereby effectively been segmented and isolated from the surrounding objects (as if cut out
by scissors).
Note that, given the reasonably long reach of the knowledge bases projecting radially outward
from the center of the primary visual layer, even objects which are interrupted by an occluding
foreground object will, in principle, have all of their visible components represented by primary
lexicons (and those coding the interrupting object(s) will be nulled). Also note that the smaller each
primary visual lexicon is (in terms of the fraction of the eyeball image it covers), the better this