Information Technology Reference
In-Depth Information
patients, who typically have damage to the right pari-
etal lobe, exhibit deficits in processing information that
appears in the left-hand side of space. For example, they
are impaired at detecting stimuli that appear in the left,
and they fail to reproduce the left-hand side of scenes
that they are asked to draw. They also have problems
that appear to be specific to the left-hand sides of objects
(Tipper & Behrmann, 1996), suggesting that parietal
spatial representations have multiple reference frames
(see also Snyder, Grieve, & Andersen, 1998).
Neural recordings from posterior parietal areas in
monkeys clearly show that these areas encode combina-
tions of both visual location and eye position (Zipser &
Andersen, 1988). Furthermore, Colby et al. (1996)
have shown that neurons in LIP that encode spatial loca-
tion information are dynamically updated to anticipate
the effects of eye movements.
The dorsal pathway may provide a spatially orga-
nized form of attention. We explore a model of this
kind of attention in section 8.5, where we see that ne-
glect can be simulated by damaging a pathway that has
spatially mapped representations. The attentional effect
produced by the spatial pathway in this model interacts
in important ways with the object-processing pathway,
enabling multiple objects to be processed without con-
fusion. Thus, although it is important to understand how
different aspects of visual information are processed in
different pathways, it is also important to understand
how these pathways interact with each other.
The correlational structure of the visual environ-
ment provides a computationally oriented way of think-
ing about why edges are represented in V1. As we
discussed in chapter 4, there are reliable correlations
among the pixels that lie along an edge of an object,
and because objects reliably tend to have such edges,
they provide the basis for a particularly useful and com-
pact representation of the environment. Once these ba-
sic pixelwise correlations are represented as edges, then
subsequent levels of processing can represent the higher
level correlations that arise from regularities in the ar-
rangement of edges (e.g., different kinds of edge inter-
sections or junctions, basic shape features, etc.), and so
on up to higher and higher levels of visual structure (as
we will see in the next model).
The objective of the first model is to show how a
V1-like network can learn to represent the correlational
structure of edges present in visual inputs received via
a simulated thalamus. Olshausen and Field (1996)
showed that a network that was presented with natu-
ral visual scenes (preprocessed in a manner generally
consistent with the contrast-enhancement properties of
the retina) could develop a realistic set of oriented edge-
detector representations. However, this network was not
based on known biological principles and was mainly
intended as a demonstration of the idea that sparse rep-
resentations provide a useful basis for encoding real-
world (visual) environments. Many other computa-
tional models of these early visual representations have
been developed, often emphasizing one or a few aspects
of the many detailed properties of V1 representations
(for recent reviews, see Swindale, 1996; Erwin, Ober-
mayer, & Schulten, 1995). These models have been
very useful in illuminating the potential relationships
between various biological and computational proper-
ties and the resulting V1 representations.
The model we present incorporates several of the
properties that have been identified in other models
as important, while using the principled, biologically
based mechanisms developed in the first part of this
text. This relatively simple model — a standard Leabra
model with one hidden layer — produces fairly realistic
V1 representations based on natural visual inputs. The
model uses the CPCA Hebbian learning algorithm on
the same preprocessed visual scenes used by Olshausen
8.3
Primary Visual Representations
The first model we explore deals with the lowest level
representations in the visual cortex (i.e., in area V1 ),
that provide the basis upon which most subsequent vi-
sual cortical processing builds. This model demon-
strates why the known properties of these representa-
tions are computationally useful given the nature of the
visual world. This demonstration emphasizes one of
the main benefits of computational models in cognitive
neuroscience — they can explain why the brain and/or
cognition has certain properties, which provides a much
deeper level of understanding than merely documenting
these properties.
Search WWH ::




Custom Search