The retina sends to the brain a parallel set of about a dozen different space-time representations of the visual world. Each of these representations is generated by a distinct set of "feature detecting" transformations. These features most likely contain all the information we need and use to analyze and interpret the visual world. They constitute a fundamental visual language that is elaborated upon at higher centers in the brain. A multi-layer CNN is presented for mimicking this new retinal model. The model is composed of several prototype 3-layer CNN units, called Complex R-units. Surfaces of activity are represented by CNN layers. Various parameter sets represent the different parts of the multi-layer retinal model. The whole model can be described by a visual language with elementary instructions of a CNN Universal Machine containing the programmable Complex R-units. Decomposition methods in time and space are discussed.