A Neural Computation Model Based on nCRF and Reverse Control Mechanism - Developing Concepts in Applied Intelligence - page 46

Each visual neuron only elicits response to the stimulation in a specific area of the
retina (or visual filed). The area is called as the classical receptive field (CRF) of the
neuron. The CRF of the GC has an antagonistic center-surround structure. With the
development of in-depth studies on human vision system, many researchers [1, 2]
found that there is still a large area outside the CRF. Stimulating such area alone fails
to elicit neural responses but can modulate the neural response to CRF stimulation.
The area is called as non-classical receptive field (nCRF). Sun C. et al. [3] found that
stimulating CRF alone can also elicit neural response when the nCRF receive large
area and appropriate stimuli. Disinhibitory nCRF helps to transmit the luminance and
grads of image local area [4].
Some models were built to stimulate disinhibitory nCRF. Li Z. et al. [5] proposed a
function with the shape of a volcano whose center is concave. Ghosh K. et al. [6]
proposed a linear function of three-Gaussian and explained low-level brightness-
contrast illusions. Qiu F. T. et al. [7] gave a model for the mechanism of mutual
inhibition within disinhibitory nCRF. These models were established by simple
reference to the mechanism of nCRF. They were just used in contrast enhancement,
edge extraction, etc. and had no top-down feedbacks. However the modulations from
high-level are very important in the neurobiology, which has been proved by
electrophysiology, anatomy and morphology.
Research shows that ganglion RF can dynamically change its size [8]. The size of
RF varies with the changes of brightness, background, length of time of stimulation,
speed of moving objects and so on. Previous nCRF models are mainly based on fixed
RF, whose dynamic characteristics are not taken into account.
Our computational model is not for segmentation. Simulating biological visual
system needs to address three fundamental issues: first, which features to
automatically choose for representation; second, what to use to achieve
representation; third, can biological nervous systems achieve this task? To make a
computational model solve these three problems simultaneously is very challenging
and far exceed the scope of the classical segmentation. By establishing ganglion cell's
dynamic receptive fields and its self-adaptive mechanism, we realized automatic
analysis and extraction of block features. By simulating the array of ganglion cells we
achieved representation. And the neural circuit we developed is supported by the
neurobiological evidence. Therefore, this paper is aimed at how to self-adaptively
extract features and form internal representations for images by a bio-inspired
structure. Limited to the length of the paper, we just presented a limited number of
experimental results.
2 Physiological Basis
2.1 Neurophysiological Mechanism of nCRF of Retinal GC
Neurophysiological studies show that [9] the horizontal cell (HC) with a wide range
of electrical synaptic connections can receive inputs from large areas of retinal
receptor cells. The HC affects the activities of receptor cells (RCs) by means of
spatial summation and feedback and then elicits the GC responses to the stimuli
within the disinhibitory nCRF through bipolar cells (BCs). Moreover, the amacrine