Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 227

hypercolumns

blobs

2−
3

4

L

R

L

R

orientations

Figure 8.5: A pinwheel structure, where orientation coding
of V1 neurons (indicated by the orientation of the lines) moves
around in a circle of neurons, with a singularity in the middle
where orientation is not clearly coded (indicated by the small
square).

occularity

Figure 8.4: Structure of a cortical hypercolumn, that repre-
sents a full range of orientations (in layers 2-3), ocular domi-
nance columns (in layer 4, one for each eye), and surface fea-
tures (in the blobs). Each such hypercolumn is focused within
one region of retinal space, and neighboring hypercolumns
represent neighboring regions.

Another interesting topographic feature of V1 is the
pinwheel (e.g., Blasdel & Salama, 1986), where all ori-
entations are represented around a circle of neurons,
with a singularity in the middle of the circle where ori-
entation is not clearly coded (figure 8.5). We will see
in the first simulation below that this pinwheel struc-
ture, along with many of the other key properties of V1
edge detector neurons and their topographic organiza-
tion, can emerge from CPCA Hebbian learning with the
kWTA activation function, together with neighborhood
interactions between neurons.

Within the interblob region, the edge detectors appear to
be organized according to orientation, such that neigh-
boring neurons encode similar orientations, with a rela-
tively smooth progression of orientation found by mov-
ing along a given direction. The ocular dominance
columns , in which V1 neurons respond preferentially
to input from one eye or the other, but not both, are
organized orthogonally to the orientation dimension.
These ocular dominance columns might be important
for stereo depth coding, although they are not present in
all mammals.
This topographic arrangement can be summarized
with the notion of a hypercolumn , that contains a
full range of orientation codings and ocular dominance
columns (i.e., one for each eye), in addition to two blobs
(figure 8.4). All the neurons within a given hypercol-
umn process roughly the same region of retinal space,
and neighboring hypercolumns process neighboring re-
gions of space. Inevitably, one must view this hyper-
column topology as more of a continuum than as the
discrete structure described here. One way of viewing
such a continuum is that the blob regions contain the
neurons tuned to a low spatial frequency (i.e., broad ar-
eas), while the interblob regions contain higher spatial
frequency cells. Nevertheless, we will see that this no-
tion of a hypercolumn is a useful abstraction for orga-
nizing models.

8.2.4 Two Visual Processing Streams

Beyond V1, visual processing appears to separate into
two major streams (figure 8.6). Initially, this split was
described in terms of a ventral “what” pathway for
processing object identity information, and a dorsal
“where” pathway for processing spatial location and
motion information (Ungerleider & Mishkin, 1982).
The ventral stream goes along the lower part of the cor-
tex from the occipital lobe to the inferior temporal cor-
tex (IT). The dorsal stream goes along the upper part of
the cortex from the occipital lobe to the parietal lobe.
The reality of what these brain areas do is probably a
bit more complex than this simple story. For example,
it is likely that the dorsal pathway plays a role in us-
ing visual information to guide motor actions, which
certainly involves spatial location information, but also
certain kinds of visual form information (Goodale &
Milner, 1992). There is clearly intercommunication be-
tween these pathways, so they cannot be considered