THE USE OF CNN MODELS IN THE SUBCORTICAL VISUAL PATHWAY

The CNN model is now a paradigm of cellular analog programmable multid imensional processor array with distributed local logic and memory. It has turned out that for biological sensory information processing thi s model is so natural that a lot of neuroanatomical structures can be directly translated into CNN models. In this paper first we show the e quivalent notions of neuroanatomy and the CNN model, motivated by stud ying the visual system. Next, various, mainly subcortical phenomena ar e studied. Simple effects like directional sensitivity and length tuni ng are modeled. A more accurate retina model has been developed taking into account some effects of amacrine cells. It is shown that the sta ndard errors occurring in simple models of retinal illusions can be el iminated by using our more accurate models including delays. LGN effec ts with and without cortical feedback are modeled as well. Their CNN m odels are simple. Furthermore, simple texture detection effects and mo tion illusions are explained by neuromorphic CNN models. Local connect ivity within a finite neighborhood, equivalent templates of different receptive field organizations, delay-templates have proved to be key i ssues. The goal is to translate known effects into CNN models and to s how a framework for further studies.