![](bg2.png)
number of decisive advantages. First, structured
nets represent visual patterns explicitly, as a two-
dimensional arrangement of local texture elements.
Second, as alluded to above, large numbers of nets
can be implemented on a comparatively narrow
neural basis in a combinatorial fashion. Third,
partial identities of different patterns are taken care
of by partial identity of the representing pieces of net
structure. Fourth, a whole hierarchy of features can
be represented in a flat structure already in primary
visual cortex (a shade of neurophysiological
evidence for the lateral connections between neurons
surfaces in the form of non-classical receptive fields,
see Allman et al. 1985). Fifth, nets that are
homeomorphic to each other (i.e., can be put into
neuron-to-neuron correspondence such that
connected neurons in one net correspond to
connected neurons in the other) can activate each
other directly, without this interaction having to be
taught, see below.
3 ACTIVATION OF NETS
Once local net structure has been established by
learning and self-interaction, the activation by visual
input takes the following form (see Fig. 1). The
sensory input selects local feature types. Each
feature type is (at least in a certain idealization)
represented redundantly by a number of neurons
with identical receptive fields. Sets of such input-
identical redundant neurons form “units”. Within a
unit there is an inhibitory system inducing winner-
take-all (WTA) dynamics (only one or a few of the
redundant neurons surviving after a short time). The
winners in this process are those neurons that form
part of a net, that is, whose activity is supported by
lateral, recurrent input.
This process of selection of the input-activated
neurons that happen to be laterally connected as a
net is an important type of implementation of
dynamic links: although the connections are actually
static, nets are dynamically activated by selection of
net-bearing neurons. For another type see below.
Local pieces of net structure can be connected
like a continuous mosaic into a larger net. This may
be compared to the image-compression scheme in
which the texture within local blocks of an image is
identified with a code-book entry (only the
identifying number of the code-book entry being
transmitted), the code-book entries tiling the image.
4 GENERATION OF NETS
Net structure in primary visual cortex is shaped by
two influences, input statistics and self-interaction.
One may assume that the genetically generated
initial structure has random short-range lateral
connections. In a first bout of organization receptive
fields of neurons are shaped by image statistics,
presumably under the influence of a sparsity
constraint (Olshausen and Field, 1996). In this
period the WTA inhibition may not yet be active,
letting neurons in a unit develop the same receptive
field. Then the network becomes sensitive to the
statistics of visual input within somewhat larger
patches (the scale being set by the range of lateral
connections) and pieces of net structure are formed
by synaptic plasticity strengthening connections
between neurons that are often co-activated and
WTA-selected, while net structure is optimized by
the interplay between (spontaneous or induced)
signal generation and Hebbian modification of
synaptic strengths under the influence of a synaptic
sparsity constraint.
5 MODALITIES
Different sub-modalities (texture, colour, depth,
motion, ..) form their own systems of net structure,
that is, representations of local patterns that are
statistically dominant in the sensory input. Each
modality is invariant to the others and has its own
local feature space structure with its own
dimensionality, three for colour, two for in-plane
motion, one for (stereo-)depth, perhaps 40 for grey-
level texture and so on. Different values along a
given feature dimension are represented by different
neurons, or rather units containing a number of
value-identical neurons. Different value-units of the
same feature dimension, forming a “column”, inhibit
each other, again in WTA fashion.
6 LATENT VARIABLES
Several units standing for different values of a sub-
modality feature may be simultaneously active to
varying degree. They may be seen as representing
different hypotheses as to the actual value of the
feature dimension. These activities thus represent
heuristic uncertainty, which during the perceptual
process needs to be reduced to certainty. In
distinction to computer graphics, realized as a
NCTA2014-InternationalConferenceonNeuralComputationTheoryandApplications
346