Neurobiological Background - Hierarchical Neural Networks for Image Interpretation - page 31

The visual system has received much attention in neurobiology and psycho-
physics. In fact, more research has been done for vision than for all other senses
together. Many details about the organization of the visual system are known at the
various levels of description. However, as of today, the function of the system has
not been understood completely.
For instance, there is debate about the neural code used by the brain. One of
the questions is whether cortical neurons are mainly driven by the average firing
rates of presynaptic neurons or by temporally coherent firing events. It is likely that
both coding schemes are used in situations where they are appropriate. In this sense,
Tsodyks and Markram [230] argue that the code depends on the rate of synaptic
depression and that a continuum between rate codes and temporal codes exists.
Another open issue is the so called binding problem. How is information about
color, motion, depth, and form, which is carried by separate neuronal pathways,
organized into cohesive perceptions of objects? Since different features of a vi-
sual scene are represented by the activity of specialized neurons that are distributed
through the visual system, all aspects of an object must be brought temporally into
association.
Treisman et al. [229] and Jualesz [114] have shown that such associations re-
quire focused attention. They found that distinctive elementary properties, such as
brightness, color, and orientation of lines, create distinctive boundaries that make
objects preattentively salient. They suggest that in a first phase of perception, all
features of the visual field are processed in parallel in a bottom-up way. In their
model a spotlight of attention highlights the features of individual objects in a serial
manner after the initial analysis. This reflects the effects of top-down attention. The
spotlight of attention requires a master map that combines details from individual
feature maps which are essential for recognition.
Another view on the effect of attention was recently proposed by Reynolds and
Desimone [190]. They assume that attention acts to increase the competitive advan-
tage of the attended stimulus so that the effect of attention is to shrink the effective
size of a neuron's receptive field around the attended stimulus, as illustrated in Fig-
ure 2.14. Now, instead of many stimuli with different characteristics, such as color
and form, only one stimulus is functionally present in the receptive field.
A different approach to the binding problem has been proposed by Singer and
Gray [216] and Eckhorn et al. [59]. They found that when an object activates a
population of neurons in the visual cortex, these neurons tend to oscillate and to
synchronize their action potentials. To bind together different visual features of the
same object, the synchrony would extend across neurons in different visual areas.
Another puzzling problem is the role of the recurrent connections, ubiquitous
in the visual system, with respect to conscious visual experience. Visual perception
is usually explained in the context of the feed-forward model of visual processing.
This model starts from the anatomical hierarchy of cortical areas, with areas V1
and V2 at the lowest levels and the inferotemporal and frontal cortex at the highest
stages. Selectivity of a neuron at a given stage is assumed to result from the or-
ganized convergence of feed-forward inputs from neurons located at lower stages.