Connectionist models applied to image compression and self-organization of the mammalian visual system.

Summary The gradient backpropagation algorithm has been used to compress images on multi-layer autoassociation networks. The solution obtained by the network after learning is realized in a very different way than the classical methods. A theoretical study on the convexity of the solution space allowed to improve the convergence speed of the network by imposing constraints on the synaptic weights. In order to improve the visual results of image compression, we introduced, in the backpropagation algorithm, the minimization of the holder norms. This algorithm has proven to be a very efficient tool and brings new solutions to the image compression problem. Our second task was to study and generalize a self-organization model to explain the formation and organization of cells characteristic of the mammalian visual system during the prenatal and postnatal periods. We propose a statistical model to study postnatal evolution. We propose a statistical model to study the postnatal evolution of the orientation columns of layer iv of the visual cortex by simulating the first visual experiences of the animal on the retina after birth (darkness, biased environments, normal visual environments). Our model takes into account recent neurobiological observations on the development of intra-cortical connections. Our simulations recover the neurobiological results showing the existence of excitatory intracortical connections between selective cells at the same orientations, and belonging to different hypercolumns.