Modeling of orientation selectivity in the visual cortex

Experimental Study and Computational Modeling of
Orientation Selectivity in the Visual Cortex

I. A. Rybak , N. A. Shevtsova, A. V. Golovan, and L. N. Podladchikova

Orientation selectivity is a remarkable property of neurons in the visual cortex which is supposed to provide the detection of local bars and edges in the processed visual images and encoding of their orientations (Hubel and Wiesel, 1962; 1974). According to the concept of columnar organization, the neighboring neurons in the visual cortex have similar orientation tunings and comprise an orientation column or iso-orientation domain (Hubel and Wiesel, 1974). A set of orientation columns with common receptive field forms a topical module of the cortex, the hypercolumn (Hubel and Wiesel, 1962; 1974).

The mechanism for orientation selectivity in the iso-orientation domain (IOD) is supposed to be based on anisotropic distribution of some fibers. The initial explanation of this mechanism, proposed by Hubel and Wiesel, was based on the anisotropic distribution of afferent connections from the inhibitory and excitatory subfields of the receptive field. Another explanation assumed the existence of anisotropic lateral excitatory connections between the neurons within the IOD (e.g. Finette et al., 1978). At the same time, Sillito (1975, 1984) found that the orientation selectivity of the cortical neurons decreased or even disappeared when intracortical inhibition was blocked. These data support the idea that the orientation sensitivity may result from an anisotropy in the lateral inhibitory connections within the orientation column (IOD).

In order to analyze this idea, we developed a computational neural model which considered IOD as a neural network with retinotopically organized afferent inputs and spatially anisotropic, reciprocal lateral inhibition provided by inhibitory interneurons.

We investigated the dynamics of neuronal responses in the IOD model to the stationary bars and edges with the "optimal" orientation (orthogonal to the direction of maximal lateral inhibition) and to the same stimuli but oriented at different angles to the "optimal" orientation (see Fig. 1).

Fig. 1. The neuronal responses in the IOD model to the "optimally" (upper row) and "non-optimally" (second and third rows) oriented bars (left column) and edges (right column).

.

Fig. 2. Histograms of responses of three types of neurons in the guinea pig visual cortex (area V1) to the optimally (upper row) and non-optimally oriented (second and third rows) light bars and a diffuse stimulus (bottom row).
(The initial and later phases of responses are marked by yellow and green respectively).

Neuron type I shows orientation selectivity only in the later phase of response
In neuron type II, both phases of response are orientation dependent
Neuron type III has a monophasic orientationally selective response

Our analysis of neuronal responses in the IOD model (Fig. 1) revealed the following:

The results of our in vivo studies (see Fig. 2) supported our model predictions: 73% of all registered orientation-selective neurons demonstrated responses whose later phases were much more sharply tuned to the stimulus orientation than their initial phases (types I and II in Fig. 2). Moreover, 43% of neurons demonstrated an orientation selectivity only in the later phases of their responses (type I in Fig. 2).

Interestingly, the orientation selectivity of the initial phase of neuron response increased with the latency of response, which may be considered as an indirect support for our model. Our explanation for this phenomenon is that neurons with more response latency start firing later, on the background of orientation-dependent inhibition provided by neurons with shorter response latencies (which, in turn, do not show orientation sensitivity in the initial phase of responses).

Our finding allows the hypothesis that the initial phases of neuron responses encode the location of visual stimuli (edges) whereas the later phases encode the stimulus orientations.

Temporal separation of information about the location of the edge and about its orientation at the neuronal level in the primary visual cortex may provide a mechanism for initial bifurcation of the visual processing into the "where" and "what" pathways (Ungerleider & Mishkin 1982) and be used for the parallel-sequential image processing under control of visual attention.

Our model of the neural network system for image processing was based on this idea. The model incorporated the IOD model as an orientation detecting element of a preprocessor. The preprocessor in turn consisted of a set of IODs tuned to the different edge orientations. Images were processed by way of fixations of an "attention window" at the sequentially selected fixation points. During image processing, the preprocessor performed the following: (i) encoding the edge location and directing the sequential shifts of the "attention window" using "where" information, extracted from the first phases of IOD neuronal responses, and (ii) encoding the edge orientation using "what" information, extracted from the second phases of IOD neuronal responses. An example of the sequential processing of the test image of "Nefertiti's head" (Fig. 4A, similar to the image used by Yarbus (1967), Fig. 3) is shown in Fig. 4B,C.

This model together with the Eckhorn model (Eckhorn et. al, 1990) provided a basis for Pulse-Coupled Neural Networks (PCNN) paradigm in Artificial Neural Networks (see T. Lindblad and J. M. Kinser. Image Processing using Pulse-Coupled Neural Networks. Springer-Verlag: 1998).

I. A. Rybak, N.A. Shevtsova, L. N. Podladchikova and V. M. Sandler (1990) Modeling of neural organization of the visual cortex and some issues of image processing by neuron-like networks. In Neural Networks: Theory and Architecture (Eds. A. Holden and V. Krukov). Manchester University Press, 117-137.

L. N. Podladchikova, I. A. Rybak, N. A. Shevtsova, and A. V. Golovan (1990) Filtration of oriented image elements and feature discrimination in the visual cortex. In Neurocomputers and Attention. Vol. 1 "Neurobiology, Synchronization and Chaos" (Eds. A.Holden and V. Krukov). Manchester University Press, 81-96.

N. A. Shevtsova, I. A. Rybak, L. N. Podladchikova, and A. V. Golovan (1991) Temporal and spatial discrimination of image features in the visual cortex. In Theoretical Aspects of Neural Networks (Eds. M. Novak and E. Pelican). World Sci. Pub. Co., 259-276.

I. A. Rybak, L. N. Podladchikova, N. A. Shevtsova, and A. V. Golovan (1991) A visual cortex domain model and its use for visual information processing. Neural Networks. 4: 3-13 <pdf>.

I. A. Rybak, N. A. Shevtsova, and V. M. Sandler (1992) A model of the neural network visual preprocessor. Neurocomputing, 4:93-102 <pdf> (The Best Paper Award of Neurocomputing in 1992).

I. A. Rybak , N. A. Shevtsova, A. V. Golovan, and L. N. Podladchikova

Fig. 1. The neuronal responses in the IOD model to the "optimally" (upper row) and "non-optimally" (second and third rows) oriented bars (left column) and edges (right column).

.

Fig. 3. Human eye movements
during image perception
(Yarbus, 1967).

Fig. 4. Model performance.

.

I. A. Rybak , N. A. Shevtsova, A. V. Golovan, and L. N. Podladchikova

Fig. 1. The neuronal responses in the IOD model to the "optimally" (upper row) and "non-optimally" (second and third rows) oriented bars (left column) and edges (right column). .

Fig. 3. Human eye movements during image perception (Yarbus, 1967).

Fig. 4. Model performance. .

Fig. 1. The neuronal responses in the IOD model to the "optimally" (upper row) and "non-optimally" (second and third rows) oriented bars (left column) and edges (right column).

.

Fig. 3. Human eye movements
during image perception
(Yarbus, 1967).

Fig. 4. Model performance.

.