| In the first part of this thesis we introduce a Bayesian approach to self-organization using prior probability assumptions about the signal as an organizing principle. This is a natural generalization of the criterion of maximizing mutual information assuming spatial coherence, as implemented by Becker and Hinton. Using our principle it is possible to self-organize Bayesian theories of vision, assuming that the priors are known, the network is capable of representing the appropriate functions and the learning algorithm converges. In a similar vein, we argue that a feedforward neural network can self-organize, by backpropagation, to approximate the maximum a posteriori (MAP) estimator of a Bayesian theory. Simulation results confirmed this for the case of image segmentation using the weak string/membrane model.; In the second part of the thesis we investigate contrast adaptation in the retinas of the larval tiger salamander and the rabbit. The function of adaptation is to match the dynamic range of the output of neuronal circuits to the dynamic range of their input. An efficient visual encoder should adapt its strategy to match the full shape of the ambient intensity distribution. We found that retinal ganglion cells, the output neurons of the retina, adapt to both image contrast--the range of ambient light intensities--and to spatial correlations within the visual scene, even at constant mean intensity. Adaptation of this type is mediated within the retinal network: two independent sites of modulation after the photoreceptor cells appear to be involved. These results demonstrate a remarkable plasticity and dynamic adaptability in retinal processing that may well contribute to contrast adaptation in the human visual system. |
