Perception and Attention - Computational Explorations in Cognitive Neuroscience - page 223

ical grounding for the models described in this chapter.
Thus, we have focused our presentation on those aspects
of visual processing that are particularly important for
object recognition, to the exclusion of many other as-
pects that are important for processing motion, color,
surface, depth, and so on. Although these other visual
properties play a role in object recognition (and atten-
tion), people are remarkably good at recognizing ob-
jects from simple line drawings, which indicates that
basic shape or form information is at least sufficient for
object recognition.
One of the key properties necessary for object recog-
nition, in addition to representing basic form informa-
tion (e.g., outline or general shape) is spatial invari-
ance , where a given object can be recognized in a rel-
atively wide range of spatial locations, sizes, rotations,
and the like. In our models, we explore location and size
invariance, which are the best documented aspects of
spatial invariance in terms of the underlying neural rep-
resentations. We will see that the visual system seems
to gradually build up this invariance over subsequent
stages of processing.

temporal

optic
chiasm

V2,V4...

nasal V1

LGN

V2,V4...

temporal

Figure 8.1: Pathway of visual information from retina to cor-
tex. Light is reflected off of surfaces, transduced in the retina,
and routed according to left and right visual fields (with the
projections from the nasal retina crossing at the optic chiasm),
to the LGN of the thalamus, which then projects up to the pri-
mary visual cortex, V1, and from there to specialized cortical
processing areas (V2, V4, and beyond).

nomena should apply more generally to other aspects of
visual processing, and to cortical processing in general.
For example, the object recognition model provides an
excellent demonstration of the power of rich, overlap-
ping distributed representations developed over many
specific training experiences to solve difficult problems.

8.2.1 The Retina

8.2 Biology of the Visual System

The retina is much more than a passive light transduc-
tion device — it provides a relatively highly processed
signal to the cortical visual areas (via the thalamus). It is
important to understand the nature of this processing be-
cause it is incorporated into our models, in some cases
by performing some appropriate form of preprocessing
on actual images, and in others by directly providing
input representations that roughly capture the effects of
retinal processing.
Retinal processing performs contrast enhancement
that emphasizes visual signal changes over space (and
time). This processing tends to enhance the edges of
objects, at the expense of coding the absolute values of
regions with relatively constant illumination (i.e., sur-
faces). One benefit of this processing is that it greatly
compresses the representation of a visual scene — in-
stead of conveying every “pixel” of illumination, fewer
and more informative edges are represented.
A complex three-layered circuit is responsible for
the retinal processing. Everything starts with the light-

The biology of the visual system begins with the eyes,
but it is not completely clear where it ends, as increas-
ingly abstract levels of visual processing merge grad-
ually into multimodal or amodal “cognitive” process-
ing. Indeed, there is considerable evidence that puta-
tively visual brain areas are activated when a person
is thinking about the semantic properties of an object
(see chapter 10). In this section, we cover the brain ar-
eas widely regarded as visual processing areas, starting
with the retina , going next to the lateral geniculate nu-
cleus (LGN) of the thalamus , then into the primary and
higher visual cortex in the occipital lobe (V1, V2, etc.),
and continuing up into the parietal and temporal lobes
(figure 8.1).
Although a vast amount is known about the visual
system, particularly the early stages of processing, we
provide only a sketch of the main findings here. Our
objective is to provide sufficient orientation and empir-