Mitsuo KAWATO, Toshio INUI, Sadayuki HONGO and Hideki HAYAKAWA
Computational Theory
and Neural Network Models of Interaction
Between Visual Cortical Areas
Abstract:We develop a computational theory and a neural network model which
coherently explains early, middle and high-level vision problems based on the
anatomical structure and physiological functions of the visual cerebral cortices.
Our computational theory is based upon a hierarchical and stochastic
model of image generation with highly redundant, multiple representations
at different description levels. We propose that feedforward neural connections
from the lower to the higher visual areas provide approximated inverses
of image generation, while feedback neural connections from the higher to
the lower areas provide forward models of image generation. We propose a
global, hierarchical model of interactions between several visual cortical areas,
in which internal representations of the 3-D world in each area are specified.
First, the solutions to several visual computational problems, such as
boundary detection, motion, color, stereo and the shape from shading problem,
are outlined in our general framework. In particular, the shape from
shading problem will be dealt with in detail by a concrete neural network
model and computer simulations. Second, brightness illusions, the Mach
band and Craik O'Brien illusion are simulated by a neural network model
based on our general framework with emphasis on the disappearance of the
illusions under high contrast conditions. Finally, a learning algorithm called
the "cross-covariance learning rule", with which the internal models of the
visual world can be acquired in the visual cerebral cortices, is proposed.