CORTICAL PERCEPTUAL CODING
Continuous Neural Field Modeling of
Primary Visual Cortex
Important unresolved questions in visual neuroscience concern the neural
circuitry by which the receptive fields of cortical cells gain their
space-time structure subservient to different functional properties.
How do simple cell receptive fields gain their Gabor-like profiles? How
do direction selectivity and spatiotemporally oriented receptive fields
originate? How do cortical complex cells gain their nonlinear response
properties?
Following a ``reverse-engineering'' approach to seek possible architectural
answers to these questions, we proposed a neural field model of simple cell
functionality. The model, based on the superposition of geniculate
(i.e., feed-forward) and intracortical (i.e., feedback) contributions,
suggests to relate the origin of highly-structured Gabor-like receptive
fields to the propagation properties of cortical activation triggered by
the specific spatial organization of the recurrent lateral inhibition
scheme adopted.
Direction selectivity and space-time oriented receptive fields emerge from
the same architectural model when recurrent inhibitory interactions occur
through spatially asymmetric schemes.
It is worthy to note that when inhibition is mediated by gating
circuits controlled by additional feed-forward signals, the resulting
visual operators present a non-linear behavior. The analysis of the
resulting operators through bi-impulsive stimulation evidenced
2nd-order Volterra kernels that closely resemble non-linear subunits of
cortical complex cells.
In conceiving these models and carrying on their analysis, major emphasis is on
the extrapolation of computational principles than on the details of
neurobiological facts. For that reason, this study
should be considered not as a realistic biological model of primary visual
cortex, but as an attempt to advance some average structural/functional
principles of intracortical connectivity and to relate them to the
underlying single-cell properties.
Back to index
Representation properties of cortical orientation maps
The efficiency of cortical representation of orientation can be evaluated
in terms of the completeness of the representation and of its redundancy
necessary to preserve information from intrinsic noise.
The analysis of the efficiency of the resulting arrangements of orientation
selective cells can be pursued by investigating how
information about any small locus in visual space affects the activity of
nearby cells on the basis of the overlap of their receptive fields.
The way in which neighboring receptive fields overlap
depends on their shape and size, and on their relative disposition
as dictated by the orientation map.
In order to have the same representation and computation capabilities
across the visual space, it is necessary to have an
uniform allocation of cells of different orientation
preference for each point (cf. the uniform coverage criterium of
Nick Swindale).
According to a simple linear model of cortical cells, we
conceived an information-theoretic framework to evaluate
coverage uniformity of several experimental orientation maps by
studying the amount and the isotropy of information transmitted
horizontally on the maps.
At each location of the cortical plane the directional
mutual information is evaluated, as a measure of coverage uniformity.
The resulting coverage uniformity map is defined by the degree of
isotropy of the mutual information.
The efficiency of an orientation map has been related to a
corresponding coverage uniformity map characterized by uniformly low
values.
This criterium applied to pinwheel orientation maps (experimental
data from the literature) reveals optimal coverage uniformity, compared to
other orientation arrangements, except at vortices or fractures.
These violations of uniformity are extremely localized and can
be selectively reduced since singular point are characterized by low
orientation tuning strength.
In perspective, the same criterium can be used (1) to verify the optimal
relationship between the average size of the
cortical point image and the average hypercolumn width;
(2) to define a self-organizing rule to obtain realistically looking
artificial orientation maps.
Back to index
last updated: 03, November 99 by silvio@dibe.unige.it