Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 245

Figure 8.16: V4 activation-based projection fields to the out-
put units.

Figure 8.15: V2 activation-based projection fields to the out-
put units, showing that each unit participates in representing
multiple objects.

Again, be sure to notice the scale shown at the bot-
tom of the window. You may want to manipulate the
scale (by typing in new numbers or pressing the control
buttons in the bottom right) to match the scale in the
grid log for the V2 receptive fields when making com-
parisons.

Question 8.5 Using the images of the objects shown
in figure 8.12 (which are in the same configuration as
the output units), explain one unit's participation in a
particular output representation based on the features
shown in its input receptive fields. (Hint: Pick a unit
that is particularly selective for specific input patterns
and specific output units, because this makes things eas-
ier to see.)

Question 8.6 (a) Based on this latest display, do V4
units appear to code for entire objects, or just parts of
different objects? Explain.

One can also compare the relative selectivity of these
V4 units for particular output units (objects) as com-
pared to the V2 units. By focusing specifically on the
number of objects a given unit clearly doesn't partici-
pate in, it should be clear that the V4 units are more
selective than the V2 units, which substantiates the idea
that the V4 units are encoding more complex combina-
tions of features that are shared by fewer objects (thus
making them more selective to particular subsets of ob-
jects). Thus, we see evidence here of the hierarchical
increase in featural complexity required to encode fea-
tural relationships while also producing spatial invari-
ance.

Press Open in the next file selection window to view
the input receptive fields for all of the V4 units.
Be sure to notice the scale shown at the bottom of the
window, that tells you how large the maximum values
are in this window. This scale shows that even though
the receptive fields appear yellow, the maximum value
represented by yellow is very small on the 0-1 scale
(around .1). Thus, each of these units is active across
a wide range of different input patterns. Furthermore,
these patterns appear to be uniformly distributed across
the input space, suggesting that in fact these units are
fully spatially invariant.

Finally, press Open once more to view the output
receptive fields for the V4 units (figure 8.16).

Iconify all the receptive field windows.