|M.Sc Student||Weltsch-Cohen Yariv|
|Subject||AGC Models for Signal Processing in the Primary Visual|
|Department||Department of Electrical and Computer Engineering||Supervisors||PROF. Ron Meir|
|PROFESSOR EMERITUS Yehoshua Zeevi|
Simple cells in the primary visual cortex are sensitive to a limited range of orientational stimuli. The spatial mapping of these cells in iso-orientation columns, as well as their temporal and in steady state responses to different stimuli, were studied in many physiological and psycho-physiological experiments. A model of a nonlinear feedback loop that exhibits Automatic Gain Control (AGC), is proposed and analyzed in the context of experimental physiological results. The model explains both temporal and orientational characteristics of simple cells in the visual cortex, and serves to show that some of the experimentally observed features are merely by-products of the main task for which it apparently exists, i.e. to provide the visual cortex with the ability to adapt itself to a wide range of intensities and contrasts of the stimuli.
Studies related to the dynamics of simple cells orientation tuning function were simulated using the AGC model, showing similar results to the physiological experiments.
A simulation of a development of a neural network, based on connections between cells in the network that incorporate the mechanism of feedback-AGC, was carried out. After 100 hebian learning iterations, with nine different stimuli in the form of light bars in each iteration, the cells became sensitive to orientation. The evolved orientation mapping resembles what has been reported with reference to physiological experiments.
The results of the study indicate that an AGC-based model provides a plausible explanation for many of the phenomena and characteristics exhibited by simple cells and cortical networks. Future research may use non-separable temporal/orientational parameters, and check the effect of lateral connections between cells in different columns. The research results presented in this thesis indicate that the model exhibits characteristics that can be utilized in image processing and image analysis. This should be productive avenue for further research.